| 1 | // icf.cc -- Identical Code Folding. |
| 2 | // |
| 3 | // Copyright 2009, 2010, 2011 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 | // This returns the buffer containing the section's contents, both |
| 217 | // text and relocs. Relocs are differentiated as those pointing to |
| 218 | // sections that could be folded and those that cannot. Only relocs |
| 219 | // pointing to sections that could be folded are recomputed on |
| 220 | // subsequent invocations of this function. |
| 221 | // Parameters : |
| 222 | // FIRST_ITERATION : true if it is the first invocation. |
| 223 | // SECN : Section for which contents are desired. |
| 224 | // SECTION_NUM : Unique section number of this section. |
| 225 | // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs |
| 226 | // to ICF sections. |
| 227 | // KEPT_SECTION_ID : Vector which maps folded sections to kept sections. |
| 228 | // SECTION_CONTENTS : Store the section's text and relocs to non-ICF |
| 229 | // sections. |
| 230 | |
| 231 | static std::string |
| 232 | get_section_contents(bool first_iteration, |
| 233 | const Section_id& secn, |
| 234 | unsigned int section_num, |
| 235 | unsigned int* num_tracked_relocs, |
| 236 | Symbol_table* symtab, |
| 237 | const std::vector<unsigned int>& kept_section_id, |
| 238 | std::vector<std::string>* section_contents) |
| 239 | { |
| 240 | // Lock the object so we can read from it. This is only called |
| 241 | // single-threaded from queue_middle_tasks, so it is OK to lock. |
| 242 | // Unfortunately we have no way to pass in a Task token. |
| 243 | const Task* dummy_task = reinterpret_cast<const Task*>(-1); |
| 244 | Task_lock_obj<Object> tl(dummy_task, secn.first); |
| 245 | |
| 246 | section_size_type plen; |
| 247 | const unsigned char* contents = NULL; |
| 248 | if (first_iteration) |
| 249 | contents = secn.first->section_contents(secn.second, &plen, false); |
| 250 | |
| 251 | // The buffer to hold all the contents including relocs. A checksum |
| 252 | // is then computed on this buffer. |
| 253 | std::string buffer; |
| 254 | std::string icf_reloc_buffer; |
| 255 | |
| 256 | if (num_tracked_relocs) |
| 257 | *num_tracked_relocs = 0; |
| 258 | |
| 259 | Icf::Reloc_info_list& reloc_info_list = |
| 260 | symtab->icf()->reloc_info_list(); |
| 261 | |
| 262 | Icf::Reloc_info_list::iterator it_reloc_info_list = |
| 263 | reloc_info_list.find(secn); |
| 264 | |
| 265 | buffer.clear(); |
| 266 | icf_reloc_buffer.clear(); |
| 267 | |
| 268 | // Process relocs and put them into the buffer. |
| 269 | |
| 270 | if (it_reloc_info_list != reloc_info_list.end()) |
| 271 | { |
| 272 | Icf::Sections_reachable_info v = |
| 273 | (it_reloc_info_list->second).section_info; |
| 274 | // Stores the information of the symbol pointed to by the reloc. |
| 275 | Icf::Symbol_info s = (it_reloc_info_list->second).symbol_info; |
| 276 | // Stores the addend and the symbol value. |
| 277 | Icf::Addend_info a = (it_reloc_info_list->second).addend_info; |
| 278 | // Stores the offset of the reloc. |
| 279 | Icf::Offset_info o = (it_reloc_info_list->second).offset_info; |
| 280 | Icf::Reloc_addend_size_info reloc_addend_size_info = |
| 281 | (it_reloc_info_list->second).reloc_addend_size_info; |
| 282 | Icf::Sections_reachable_info::iterator it_v = v.begin(); |
| 283 | Icf::Symbol_info::iterator it_s = s.begin(); |
| 284 | Icf::Addend_info::iterator it_a = a.begin(); |
| 285 | Icf::Offset_info::iterator it_o = o.begin(); |
| 286 | Icf::Reloc_addend_size_info::iterator it_addend_size = |
| 287 | reloc_addend_size_info.begin(); |
| 288 | |
| 289 | for (; it_v != v.end(); ++it_v, ++it_s, ++it_a, ++it_o, ++it_addend_size) |
| 290 | { |
| 291 | if (first_iteration |
| 292 | && it_v->first != NULL) |
| 293 | { |
| 294 | Symbol_location loc; |
| 295 | loc.object = it_v->first; |
| 296 | loc.shndx = it_v->second; |
| 297 | loc.offset = convert_types<off_t, long long>(it_a->first |
| 298 | + it_a->second); |
| 299 | // Look through function descriptors |
| 300 | parameters->target().function_location(&loc); |
| 301 | if (loc.shndx != it_v->second) |
| 302 | { |
| 303 | it_v->second = loc.shndx; |
| 304 | // Modify symvalue/addend to the code entry. |
| 305 | it_a->first = loc.offset; |
| 306 | it_a->second = 0; |
| 307 | } |
| 308 | } |
| 309 | |
| 310 | // ADDEND_STR stores the symbol value and addend and offset, |
| 311 | // each at most 16 hex digits long. it_a points to a pair |
| 312 | // where first is the symbol value and second is the |
| 313 | // addend. |
| 314 | char addend_str[50]; |
| 315 | |
| 316 | // It would be nice if we could use format macros in inttypes.h |
| 317 | // here but there are not in ISO/IEC C++ 1998. |
| 318 | snprintf(addend_str, sizeof(addend_str), "%llx %llx %llux", |
| 319 | static_cast<long long>((*it_a).first), |
| 320 | static_cast<long long>((*it_a).second), |
| 321 | static_cast<unsigned long long>(*it_o)); |
| 322 | |
| 323 | // If the symbol pointed to by the reloc is not in an ordinary |
| 324 | // section or if the symbol type is not FROM_OBJECT, then the |
| 325 | // object is NULL. |
| 326 | if (it_v->first == NULL) |
| 327 | { |
| 328 | if (first_iteration) |
| 329 | { |
| 330 | // If the symbol name is available, use it. |
| 331 | if ((*it_s) != NULL) |
| 332 | buffer.append((*it_s)->name()); |
| 333 | // Append the addend. |
| 334 | buffer.append(addend_str); |
| 335 | buffer.append("@"); |
| 336 | } |
| 337 | continue; |
| 338 | } |
| 339 | |
| 340 | Section_id reloc_secn(it_v->first, it_v->second); |
| 341 | |
| 342 | // If this reloc turns back and points to the same section, |
| 343 | // like a recursive call, use a special symbol to mark this. |
| 344 | if (reloc_secn.first == secn.first |
| 345 | && reloc_secn.second == secn.second) |
| 346 | { |
| 347 | if (first_iteration) |
| 348 | { |
| 349 | buffer.append("R"); |
| 350 | buffer.append(addend_str); |
| 351 | buffer.append("@"); |
| 352 | } |
| 353 | continue; |
| 354 | } |
| 355 | Icf::Uniq_secn_id_map& section_id_map = |
| 356 | symtab->icf()->section_to_int_map(); |
| 357 | Icf::Uniq_secn_id_map::iterator section_id_map_it = |
| 358 | section_id_map.find(reloc_secn); |
| 359 | bool is_sym_preemptible = (*it_s != NULL |
| 360 | && !(*it_s)->is_from_dynobj() |
| 361 | && !(*it_s)->is_undefined() |
| 362 | && (*it_s)->is_preemptible()); |
| 363 | if (!is_sym_preemptible |
| 364 | && section_id_map_it != section_id_map.end()) |
| 365 | { |
| 366 | // This is a reloc to a section that might be folded. |
| 367 | if (num_tracked_relocs) |
| 368 | (*num_tracked_relocs)++; |
| 369 | |
| 370 | char kept_section_str[10]; |
| 371 | unsigned int secn_id = section_id_map_it->second; |
| 372 | snprintf(kept_section_str, sizeof(kept_section_str), "%u", |
| 373 | kept_section_id[secn_id]); |
| 374 | if (first_iteration) |
| 375 | { |
| 376 | buffer.append("ICF_R"); |
| 377 | buffer.append(addend_str); |
| 378 | } |
| 379 | icf_reloc_buffer.append(kept_section_str); |
| 380 | // Append the addend. |
| 381 | icf_reloc_buffer.append(addend_str); |
| 382 | icf_reloc_buffer.append("@"); |
| 383 | } |
| 384 | else |
| 385 | { |
| 386 | // This is a reloc to a section that cannot be folded. |
| 387 | // Process it only in the first iteration. |
| 388 | if (!first_iteration) |
| 389 | continue; |
| 390 | |
| 391 | uint64_t secn_flags = (it_v->first)->section_flags(it_v->second); |
| 392 | // This reloc points to a merge section. Hash the |
| 393 | // contents of this section. |
| 394 | if ((secn_flags & elfcpp::SHF_MERGE) != 0 |
| 395 | && parameters->target().can_icf_inline_merge_sections()) |
| 396 | { |
| 397 | uint64_t entsize = |
| 398 | (it_v->first)->section_entsize(it_v->second); |
| 399 | long long offset = it_a->first; |
| 400 | |
| 401 | unsigned long long addend = it_a->second; |
| 402 | // Ignoring the addend when it is a negative value. See the |
| 403 | // comments in Merged_symbol_value::Value in object.h. |
| 404 | if (addend < 0xffffff00) |
| 405 | offset = offset + addend; |
| 406 | |
| 407 | // For SHT_REL relocation sections, the addend is stored in the |
| 408 | // text section at the relocation offset. |
| 409 | uint64_t reloc_addend_value = 0; |
| 410 | const unsigned char* reloc_addend_ptr = |
| 411 | contents + static_cast<unsigned long long>(*it_o); |
| 412 | switch(*it_addend_size) |
| 413 | { |
| 414 | case 0: |
| 415 | { |
| 416 | break; |
| 417 | } |
| 418 | case 1: |
| 419 | { |
| 420 | reloc_addend_value = |
| 421 | read_from_pointer<8>(reloc_addend_ptr); |
| 422 | break; |
| 423 | } |
| 424 | case 2: |
| 425 | { |
| 426 | reloc_addend_value = |
| 427 | read_from_pointer<16>(reloc_addend_ptr); |
| 428 | break; |
| 429 | } |
| 430 | case 4: |
| 431 | { |
| 432 | reloc_addend_value = |
| 433 | read_from_pointer<32>(reloc_addend_ptr); |
| 434 | break; |
| 435 | } |
| 436 | case 8: |
| 437 | { |
| 438 | reloc_addend_value = |
| 439 | read_from_pointer<64>(reloc_addend_ptr); |
| 440 | break; |
| 441 | } |
| 442 | default: |
| 443 | gold_unreachable(); |
| 444 | } |
| 445 | offset = offset + reloc_addend_value; |
| 446 | |
| 447 | section_size_type secn_len; |
| 448 | const unsigned char* str_contents = |
| 449 | (it_v->first)->section_contents(it_v->second, |
| 450 | &secn_len, |
| 451 | false) + offset; |
| 452 | if ((secn_flags & elfcpp::SHF_STRINGS) != 0) |
| 453 | { |
| 454 | // String merge section. |
| 455 | const char* str_char = |
| 456 | reinterpret_cast<const char*>(str_contents); |
| 457 | switch(entsize) |
| 458 | { |
| 459 | case 1: |
| 460 | { |
| 461 | buffer.append(str_char); |
| 462 | break; |
| 463 | } |
| 464 | case 2: |
| 465 | { |
| 466 | const uint16_t* ptr_16 = |
| 467 | reinterpret_cast<const uint16_t*>(str_char); |
| 468 | unsigned int strlen_16 = 0; |
| 469 | // Find the NULL character. |
| 470 | while(*(ptr_16 + strlen_16) != 0) |
| 471 | strlen_16++; |
| 472 | buffer.append(str_char, strlen_16 * 2); |
| 473 | } |
| 474 | break; |
| 475 | case 4: |
| 476 | { |
| 477 | const uint32_t* ptr_32 = |
| 478 | reinterpret_cast<const uint32_t*>(str_char); |
| 479 | unsigned int strlen_32 = 0; |
| 480 | // Find the NULL character. |
| 481 | while(*(ptr_32 + strlen_32) != 0) |
| 482 | strlen_32++; |
| 483 | buffer.append(str_char, strlen_32 * 4); |
| 484 | } |
| 485 | break; |
| 486 | default: |
| 487 | gold_unreachable(); |
| 488 | } |
| 489 | } |
| 490 | else |
| 491 | { |
| 492 | // Use the entsize to determine the length. |
| 493 | buffer.append(reinterpret_cast<const |
| 494 | char*>(str_contents), |
| 495 | entsize); |
| 496 | } |
| 497 | buffer.append("@"); |
| 498 | } |
| 499 | else if ((*it_s) != NULL) |
| 500 | { |
| 501 | // If symbol name is available use that. |
| 502 | buffer.append((*it_s)->name()); |
| 503 | // Append the addend. |
| 504 | buffer.append(addend_str); |
| 505 | buffer.append("@"); |
| 506 | } |
| 507 | else |
| 508 | { |
| 509 | // Symbol name is not available, like for a local symbol, |
| 510 | // use object and section id. |
| 511 | buffer.append(it_v->first->name()); |
| 512 | char secn_id[10]; |
| 513 | snprintf(secn_id, sizeof(secn_id), "%u",it_v->second); |
| 514 | buffer.append(secn_id); |
| 515 | // Append the addend. |
| 516 | buffer.append(addend_str); |
| 517 | buffer.append("@"); |
| 518 | } |
| 519 | } |
| 520 | } |
| 521 | } |
| 522 | |
| 523 | if (first_iteration) |
| 524 | { |
| 525 | buffer.append("Contents = "); |
| 526 | buffer.append(reinterpret_cast<const char*>(contents), plen); |
| 527 | // Store the section contents that dont change to avoid recomputing |
| 528 | // during the next call to this function. |
| 529 | (*section_contents)[section_num] = buffer; |
| 530 | } |
| 531 | else |
| 532 | { |
| 533 | gold_assert(buffer.empty()); |
| 534 | // Reuse the contents computed in the previous iteration. |
| 535 | buffer.append((*section_contents)[section_num]); |
| 536 | } |
| 537 | |
| 538 | buffer.append(icf_reloc_buffer); |
| 539 | return buffer; |
| 540 | } |
| 541 | |
| 542 | // This function computes a checksum on each section to detect and form |
| 543 | // groups of identical sections. The first iteration does this for all |
| 544 | // sections. |
| 545 | // Further iterations do this only for the kept sections from each group to |
| 546 | // determine if larger groups of identical sections could be formed. The |
| 547 | // first section in each group is the kept section for that group. |
| 548 | // |
| 549 | // CRC32 is the checksumming algorithm and can have collisions. That is, |
| 550 | // two sections with different contents can have the same checksum. Hence, |
| 551 | // a multimap is used to maintain more than one group of checksum |
| 552 | // identical sections. A section is added to a group only after its |
| 553 | // contents are explicitly compared with the kept section of the group. |
| 554 | // |
| 555 | // Parameters : |
| 556 | // ITERATION_NUM : Invocation instance of this function. |
| 557 | // NUM_TRACKED_RELOCS : Vector reference to store the number of relocs |
| 558 | // to ICF sections. |
| 559 | // KEPT_SECTION_ID : Vector which maps folded sections to kept sections. |
| 560 | // ID_SECTION : Vector mapping a section to an unique integer. |
| 561 | // IS_SECN_OR_GROUP_UNIQUE : To check if a section or a group of identical |
| 562 | // sections is already known to be unique. |
| 563 | // SECTION_CONTENTS : Store the section's text and relocs to non-ICF |
| 564 | // sections. |
| 565 | |
| 566 | static bool |
| 567 | match_sections(unsigned int iteration_num, |
| 568 | Symbol_table* symtab, |
| 569 | std::vector<unsigned int>* num_tracked_relocs, |
| 570 | std::vector<unsigned int>* kept_section_id, |
| 571 | const std::vector<Section_id>& id_section, |
| 572 | std::vector<bool>* is_secn_or_group_unique, |
| 573 | std::vector<std::string>* section_contents) |
| 574 | { |
| 575 | Unordered_multimap<uint32_t, unsigned int> section_cksum; |
| 576 | std::pair<Unordered_multimap<uint32_t, unsigned int>::iterator, |
| 577 | Unordered_multimap<uint32_t, unsigned int>::iterator> key_range; |
| 578 | bool converged = true; |
| 579 | |
| 580 | if (iteration_num == 1) |
| 581 | preprocess_for_unique_sections(id_section, |
| 582 | is_secn_or_group_unique, |
| 583 | NULL); |
| 584 | else |
| 585 | preprocess_for_unique_sections(id_section, |
| 586 | is_secn_or_group_unique, |
| 587 | section_contents); |
| 588 | |
| 589 | std::vector<std::string> full_section_contents; |
| 590 | |
| 591 | for (unsigned int i = 0; i < id_section.size(); i++) |
| 592 | { |
| 593 | full_section_contents.push_back(""); |
| 594 | if ((*is_secn_or_group_unique)[i]) |
| 595 | continue; |
| 596 | |
| 597 | Section_id secn = id_section[i]; |
| 598 | std::string this_secn_contents; |
| 599 | uint32_t cksum; |
| 600 | if (iteration_num == 1) |
| 601 | { |
| 602 | unsigned int num_relocs = 0; |
| 603 | this_secn_contents = get_section_contents(true, secn, i, &num_relocs, |
| 604 | symtab, (*kept_section_id), |
| 605 | section_contents); |
| 606 | (*num_tracked_relocs)[i] = num_relocs; |
| 607 | } |
| 608 | else |
| 609 | { |
| 610 | if ((*kept_section_id)[i] != i) |
| 611 | { |
| 612 | // This section is already folded into something. See |
| 613 | // if it should point to a different kept section. |
| 614 | unsigned int kept_section = (*kept_section_id)[i]; |
| 615 | if (kept_section != (*kept_section_id)[kept_section]) |
| 616 | { |
| 617 | (*kept_section_id)[i] = (*kept_section_id)[kept_section]; |
| 618 | } |
| 619 | continue; |
| 620 | } |
| 621 | this_secn_contents = get_section_contents(false, secn, i, NULL, |
| 622 | symtab, (*kept_section_id), |
| 623 | section_contents); |
| 624 | } |
| 625 | |
| 626 | const unsigned char* this_secn_contents_array = |
| 627 | reinterpret_cast<const unsigned char*>(this_secn_contents.c_str()); |
| 628 | cksum = xcrc32(this_secn_contents_array, this_secn_contents.length(), |
| 629 | 0xffffffff); |
| 630 | size_t count = section_cksum.count(cksum); |
| 631 | |
| 632 | if (count == 0) |
| 633 | { |
| 634 | // Start a group with this cksum. |
| 635 | section_cksum.insert(std::make_pair(cksum, i)); |
| 636 | full_section_contents[i] = this_secn_contents; |
| 637 | } |
| 638 | else |
| 639 | { |
| 640 | key_range = section_cksum.equal_range(cksum); |
| 641 | Unordered_multimap<uint32_t, unsigned int>::iterator it; |
| 642 | // Search all the groups with this cksum for a match. |
| 643 | for (it = key_range.first; it != key_range.second; ++it) |
| 644 | { |
| 645 | unsigned int kept_section = it->second; |
| 646 | if (full_section_contents[kept_section].length() |
| 647 | != this_secn_contents.length()) |
| 648 | continue; |
| 649 | if (memcmp(full_section_contents[kept_section].c_str(), |
| 650 | this_secn_contents.c_str(), |
| 651 | this_secn_contents.length()) != 0) |
| 652 | continue; |
| 653 | (*kept_section_id)[i] = kept_section; |
| 654 | converged = false; |
| 655 | break; |
| 656 | } |
| 657 | if (it == key_range.second) |
| 658 | { |
| 659 | // Create a new group for this cksum. |
| 660 | section_cksum.insert(std::make_pair(cksum, i)); |
| 661 | full_section_contents[i] = this_secn_contents; |
| 662 | } |
| 663 | } |
| 664 | // If there are no relocs to foldable sections do not process |
| 665 | // this section any further. |
| 666 | if (iteration_num == 1 && (*num_tracked_relocs)[i] == 0) |
| 667 | (*is_secn_or_group_unique)[i] = true; |
| 668 | } |
| 669 | |
| 670 | return converged; |
| 671 | } |
| 672 | |
| 673 | // During safe icf (--icf=safe), only fold functions that are ctors or dtors. |
| 674 | // This function returns true if the section name is that of a ctor or a dtor. |
| 675 | |
| 676 | static bool |
| 677 | is_function_ctor_or_dtor(const std::string& section_name) |
| 678 | { |
| 679 | const char* mangled_func_name = strrchr(section_name.c_str(), '.'); |
| 680 | gold_assert(mangled_func_name != NULL); |
| 681 | if ((is_prefix_of("._ZN", mangled_func_name) |
| 682 | || is_prefix_of("._ZZ", mangled_func_name)) |
| 683 | && (is_gnu_v3_mangled_ctor(mangled_func_name + 1) |
| 684 | || is_gnu_v3_mangled_dtor(mangled_func_name + 1))) |
| 685 | { |
| 686 | return true; |
| 687 | } |
| 688 | return false; |
| 689 | } |
| 690 | |
| 691 | // This is the main ICF function called in gold.cc. This does the |
| 692 | // initialization and calls match_sections repeatedly (twice by default) |
| 693 | // which computes the crc checksums and detects identical functions. |
| 694 | |
| 695 | void |
| 696 | Icf::find_identical_sections(const Input_objects* input_objects, |
| 697 | Symbol_table* symtab) |
| 698 | { |
| 699 | unsigned int section_num = 0; |
| 700 | std::vector<unsigned int> num_tracked_relocs; |
| 701 | std::vector<bool> is_secn_or_group_unique; |
| 702 | std::vector<std::string> section_contents; |
| 703 | const Target& target = parameters->target(); |
| 704 | |
| 705 | // Decide which sections are possible candidates first. |
| 706 | |
| 707 | for (Input_objects::Relobj_iterator p = input_objects->relobj_begin(); |
| 708 | p != input_objects->relobj_end(); |
| 709 | ++p) |
| 710 | { |
| 711 | // Lock the object so we can read from it. This is only called |
| 712 | // single-threaded from queue_middle_tasks, so it is OK to lock. |
| 713 | // Unfortunately we have no way to pass in a Task token. |
| 714 | const Task* dummy_task = reinterpret_cast<const Task*>(-1); |
| 715 | Task_lock_obj<Object> tl(dummy_task, *p); |
| 716 | |
| 717 | for (unsigned int i = 0;i < (*p)->shnum(); ++i) |
| 718 | { |
| 719 | const std::string section_name = (*p)->section_name(i); |
| 720 | if (!is_section_foldable_candidate(section_name)) |
| 721 | continue; |
| 722 | if (!(*p)->is_section_included(i)) |
| 723 | continue; |
| 724 | if (parameters->options().gc_sections() |
| 725 | && symtab->gc()->is_section_garbage(*p, i)) |
| 726 | continue; |
| 727 | // With --icf=safe, check if the mangled function name is a ctor |
| 728 | // or a dtor. The mangled function name can be obtained from the |
| 729 | // section name by stripping the section prefix. |
| 730 | if (parameters->options().icf_safe_folding() |
| 731 | && !is_function_ctor_or_dtor(section_name) |
| 732 | && (!target.can_check_for_function_pointers() |
| 733 | || section_has_function_pointers(*p, i))) |
| 734 | { |
| 735 | continue; |
| 736 | } |
| 737 | this->id_section_.push_back(Section_id(*p, i)); |
| 738 | this->section_id_[Section_id(*p, i)] = section_num; |
| 739 | this->kept_section_id_.push_back(section_num); |
| 740 | num_tracked_relocs.push_back(0); |
| 741 | is_secn_or_group_unique.push_back(false); |
| 742 | section_contents.push_back(""); |
| 743 | section_num++; |
| 744 | } |
| 745 | } |
| 746 | |
| 747 | unsigned int num_iterations = 0; |
| 748 | |
| 749 | // Default number of iterations to run ICF is 2. |
| 750 | unsigned int max_iterations = (parameters->options().icf_iterations() > 0) |
| 751 | ? parameters->options().icf_iterations() |
| 752 | : 2; |
| 753 | |
| 754 | bool converged = false; |
| 755 | |
| 756 | while (!converged && (num_iterations < max_iterations)) |
| 757 | { |
| 758 | num_iterations++; |
| 759 | converged = match_sections(num_iterations, symtab, |
| 760 | &num_tracked_relocs, &this->kept_section_id_, |
| 761 | this->id_section_, &is_secn_or_group_unique, |
| 762 | §ion_contents); |
| 763 | } |
| 764 | |
| 765 | if (parameters->options().print_icf_sections()) |
| 766 | { |
| 767 | if (converged) |
| 768 | gold_info(_("%s: ICF Converged after %u iteration(s)"), |
| 769 | program_name, num_iterations); |
| 770 | else |
| 771 | gold_info(_("%s: ICF stopped after %u iteration(s)"), |
| 772 | program_name, num_iterations); |
| 773 | } |
| 774 | |
| 775 | // Unfold --keep-unique symbols. |
| 776 | for (options::String_set::const_iterator p = |
| 777 | parameters->options().keep_unique_begin(); |
| 778 | p != parameters->options().keep_unique_end(); |
| 779 | ++p) |
| 780 | { |
| 781 | const char* name = p->c_str(); |
| 782 | Symbol* sym = symtab->lookup(name); |
| 783 | if (sym == NULL) |
| 784 | { |
| 785 | gold_warning(_("Could not find symbol %s to unfold\n"), name); |
| 786 | } |
| 787 | else if (sym->source() == Symbol::FROM_OBJECT |
| 788 | && !sym->object()->is_dynamic()) |
| 789 | { |
| 790 | Object* obj = sym->object(); |
| 791 | bool is_ordinary; |
| 792 | unsigned int shndx = sym->shndx(&is_ordinary); |
| 793 | if (is_ordinary) |
| 794 | { |
| 795 | this->unfold_section(obj, shndx); |
| 796 | } |
| 797 | } |
| 798 | |
| 799 | } |
| 800 | |
| 801 | this->icf_ready(); |
| 802 | } |
| 803 | |
| 804 | // Unfolds the section denoted by OBJ and SHNDX if folded. |
| 805 | |
| 806 | void |
| 807 | Icf::unfold_section(Object* obj, unsigned int shndx) |
| 808 | { |
| 809 | Section_id secn(obj, shndx); |
| 810 | Uniq_secn_id_map::iterator it = this->section_id_.find(secn); |
| 811 | if (it == this->section_id_.end()) |
| 812 | return; |
| 813 | unsigned int section_num = it->second; |
| 814 | unsigned int kept_section_id = this->kept_section_id_[section_num]; |
| 815 | if (kept_section_id != section_num) |
| 816 | this->kept_section_id_[section_num] = section_num; |
| 817 | } |
| 818 | |
| 819 | // This function determines if the section corresponding to the |
| 820 | // given object and index is folded based on if the kept section |
| 821 | // is different from this section. |
| 822 | |
| 823 | bool |
| 824 | Icf::is_section_folded(Object* obj, unsigned int shndx) |
| 825 | { |
| 826 | Section_id secn(obj, shndx); |
| 827 | Uniq_secn_id_map::iterator it = this->section_id_.find(secn); |
| 828 | if (it == this->section_id_.end()) |
| 829 | return false; |
| 830 | unsigned int section_num = it->second; |
| 831 | unsigned int kept_section_id = this->kept_section_id_[section_num]; |
| 832 | return kept_section_id != section_num; |
| 833 | } |
| 834 | |
| 835 | // This function returns the folded section for the given section. |
| 836 | |
| 837 | Section_id |
| 838 | Icf::get_folded_section(Object* dup_obj, unsigned int dup_shndx) |
| 839 | { |
| 840 | Section_id dup_secn(dup_obj, dup_shndx); |
| 841 | Uniq_secn_id_map::iterator it = this->section_id_.find(dup_secn); |
| 842 | gold_assert(it != this->section_id_.end()); |
| 843 | unsigned int section_num = it->second; |
| 844 | unsigned int kept_section_id = this->kept_section_id_[section_num]; |
| 845 | Section_id folded_section = this->id_section_[kept_section_id]; |
| 846 | return folded_section; |
| 847 | } |
| 848 | |
| 849 | } // End of namespace gold. |