| 1 | // dwarf_reader.cc -- parse dwarf2/3 debug information |
| 2 | |
| 3 | // Copyright 2007, 2008 Free Software Foundation, Inc. |
| 4 | // Written by Ian Lance Taylor <iant@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 | #include "gold.h" |
| 24 | |
| 25 | #include <algorithm> |
| 26 | #include <vector> |
| 27 | |
| 28 | #include "elfcpp_swap.h" |
| 29 | #include "dwarf.h" |
| 30 | #include "object.h" |
| 31 | #include "parameters.h" |
| 32 | #include "reloc.h" |
| 33 | #include "dwarf_reader.h" |
| 34 | |
| 35 | namespace { |
| 36 | |
| 37 | // Read an unsigned LEB128 number. Each byte contains 7 bits of |
| 38 | // information, plus one bit saying whether the number continues or |
| 39 | // not. |
| 40 | |
| 41 | uint64_t |
| 42 | read_unsigned_LEB_128(const unsigned char* buffer, size_t* len) |
| 43 | { |
| 44 | uint64_t result = 0; |
| 45 | size_t num_read = 0; |
| 46 | unsigned int shift = 0; |
| 47 | unsigned char byte; |
| 48 | |
| 49 | do |
| 50 | { |
| 51 | byte = *buffer++; |
| 52 | num_read++; |
| 53 | result |= (static_cast<uint64_t>(byte & 0x7f)) << shift; |
| 54 | shift += 7; |
| 55 | } |
| 56 | while (byte & 0x80); |
| 57 | |
| 58 | *len = num_read; |
| 59 | |
| 60 | return result; |
| 61 | } |
| 62 | |
| 63 | // Read a signed LEB128 number. These are like regular LEB128 |
| 64 | // numbers, except the last byte may have a sign bit set. |
| 65 | |
| 66 | int64_t |
| 67 | read_signed_LEB_128(const unsigned char* buffer, size_t* len) |
| 68 | { |
| 69 | int64_t result = 0; |
| 70 | int shift = 0; |
| 71 | size_t num_read = 0; |
| 72 | unsigned char byte; |
| 73 | |
| 74 | do |
| 75 | { |
| 76 | byte = *buffer++; |
| 77 | num_read++; |
| 78 | result |= (static_cast<uint64_t>(byte & 0x7f) << shift); |
| 79 | shift += 7; |
| 80 | } |
| 81 | while (byte & 0x80); |
| 82 | |
| 83 | if ((shift < 8 * static_cast<int>(sizeof(result))) && (byte & 0x40)) |
| 84 | result |= -((static_cast<int64_t>(1)) << shift); |
| 85 | *len = num_read; |
| 86 | return result; |
| 87 | } |
| 88 | |
| 89 | } // End anonymous namespace. |
| 90 | |
| 91 | |
| 92 | namespace gold { |
| 93 | |
| 94 | // This is the format of a DWARF2/3 line state machine that we process |
| 95 | // opcodes using. There is no need for anything outside the lineinfo |
| 96 | // processor to know how this works. |
| 97 | |
| 98 | struct LineStateMachine |
| 99 | { |
| 100 | int file_num; |
| 101 | uint64_t address; |
| 102 | int line_num; |
| 103 | int column_num; |
| 104 | unsigned int shndx; // the section address refers to |
| 105 | bool is_stmt; // stmt means statement. |
| 106 | bool basic_block; |
| 107 | bool end_sequence; |
| 108 | }; |
| 109 | |
| 110 | static void |
| 111 | ResetLineStateMachine(struct LineStateMachine* lsm, bool default_is_stmt) |
| 112 | { |
| 113 | lsm->file_num = 1; |
| 114 | lsm->address = 0; |
| 115 | lsm->line_num = 1; |
| 116 | lsm->column_num = 0; |
| 117 | lsm->shndx = -1U; |
| 118 | lsm->is_stmt = default_is_stmt; |
| 119 | lsm->basic_block = false; |
| 120 | lsm->end_sequence = false; |
| 121 | } |
| 122 | |
| 123 | template<int size, bool big_endian> |
| 124 | Sized_dwarf_line_info<size, big_endian>::Sized_dwarf_line_info(Object* object, |
| 125 | off_t read_shndx) |
| 126 | : data_valid_(false), buffer_(NULL), symtab_buffer_(NULL), |
| 127 | directories_(), files_(), current_header_index_(-1) |
| 128 | { |
| 129 | unsigned int debug_shndx; |
| 130 | for (debug_shndx = 0; debug_shndx < object->shnum(); ++debug_shndx) |
| 131 | // FIXME: do this more efficiently: section_name() isn't super-fast |
| 132 | if (object->section_name(debug_shndx) == ".debug_line") |
| 133 | { |
| 134 | section_size_type buffer_size; |
| 135 | this->buffer_ = object->section_contents(debug_shndx, &buffer_size, |
| 136 | false); |
| 137 | this->buffer_end_ = this->buffer_ + buffer_size; |
| 138 | break; |
| 139 | } |
| 140 | if (this->buffer_ == NULL) |
| 141 | return; |
| 142 | |
| 143 | // Find the relocation section for ".debug_line". |
| 144 | // We expect these for relobjs (.o's) but not dynobjs (.so's). |
| 145 | bool got_relocs = false; |
| 146 | for (unsigned int reloc_shndx = 0; |
| 147 | reloc_shndx < object->shnum(); |
| 148 | ++reloc_shndx) |
| 149 | { |
| 150 | unsigned int reloc_sh_type = object->section_type(reloc_shndx); |
| 151 | if ((reloc_sh_type == elfcpp::SHT_REL |
| 152 | || reloc_sh_type == elfcpp::SHT_RELA) |
| 153 | && object->section_info(reloc_shndx) == debug_shndx) |
| 154 | { |
| 155 | got_relocs = this->track_relocs_.initialize(object, reloc_shndx, |
| 156 | reloc_sh_type); |
| 157 | break; |
| 158 | } |
| 159 | } |
| 160 | |
| 161 | // Finally, we need the symtab section to interpret the relocs. |
| 162 | if (got_relocs) |
| 163 | { |
| 164 | unsigned int symtab_shndx; |
| 165 | for (symtab_shndx = 0; symtab_shndx < object->shnum(); ++symtab_shndx) |
| 166 | if (object->section_type(symtab_shndx) == elfcpp::SHT_SYMTAB) |
| 167 | { |
| 168 | this->symtab_buffer_ = object->section_contents( |
| 169 | symtab_shndx, &this->symtab_buffer_size_, false); |
| 170 | break; |
| 171 | } |
| 172 | if (this->symtab_buffer_ == NULL) |
| 173 | return; |
| 174 | } |
| 175 | |
| 176 | // Now that we have successfully read all the data, parse the debug |
| 177 | // info. |
| 178 | this->data_valid_ = true; |
| 179 | this->read_line_mappings(object, read_shndx); |
| 180 | } |
| 181 | |
| 182 | // Read the DWARF header. |
| 183 | |
| 184 | template<int size, bool big_endian> |
| 185 | const unsigned char* |
| 186 | Sized_dwarf_line_info<size, big_endian>::read_header_prolog( |
| 187 | const unsigned char* lineptr) |
| 188 | { |
| 189 | uint32_t initial_length = elfcpp::Swap_unaligned<32, big_endian>::readval(lineptr); |
| 190 | lineptr += 4; |
| 191 | |
| 192 | // In DWARF2/3, if the initial length is all 1 bits, then the offset |
| 193 | // size is 8 and we need to read the next 8 bytes for the real length. |
| 194 | if (initial_length == 0xffffffff) |
| 195 | { |
| 196 | header_.offset_size = 8; |
| 197 | initial_length = elfcpp::Swap_unaligned<64, big_endian>::readval(lineptr); |
| 198 | lineptr += 8; |
| 199 | } |
| 200 | else |
| 201 | header_.offset_size = 4; |
| 202 | |
| 203 | header_.total_length = initial_length; |
| 204 | |
| 205 | gold_assert(lineptr + header_.total_length <= buffer_end_); |
| 206 | |
| 207 | header_.version = elfcpp::Swap_unaligned<16, big_endian>::readval(lineptr); |
| 208 | lineptr += 2; |
| 209 | |
| 210 | if (header_.offset_size == 4) |
| 211 | header_.prologue_length = elfcpp::Swap_unaligned<32, big_endian>::readval(lineptr); |
| 212 | else |
| 213 | header_.prologue_length = elfcpp::Swap_unaligned<64, big_endian>::readval(lineptr); |
| 214 | lineptr += header_.offset_size; |
| 215 | |
| 216 | header_.min_insn_length = *lineptr; |
| 217 | lineptr += 1; |
| 218 | |
| 219 | header_.default_is_stmt = *lineptr; |
| 220 | lineptr += 1; |
| 221 | |
| 222 | header_.line_base = *reinterpret_cast<const signed char*>(lineptr); |
| 223 | lineptr += 1; |
| 224 | |
| 225 | header_.line_range = *lineptr; |
| 226 | lineptr += 1; |
| 227 | |
| 228 | header_.opcode_base = *lineptr; |
| 229 | lineptr += 1; |
| 230 | |
| 231 | header_.std_opcode_lengths.reserve(header_.opcode_base + 1); |
| 232 | header_.std_opcode_lengths[0] = 0; |
| 233 | for (int i = 1; i < header_.opcode_base; i++) |
| 234 | { |
| 235 | header_.std_opcode_lengths[i] = *lineptr; |
| 236 | lineptr += 1; |
| 237 | } |
| 238 | |
| 239 | return lineptr; |
| 240 | } |
| 241 | |
| 242 | // The header for a debug_line section is mildly complicated, because |
| 243 | // the line info is very tightly encoded. |
| 244 | |
| 245 | template<int size, bool big_endian> |
| 246 | const unsigned char* |
| 247 | Sized_dwarf_line_info<size, big_endian>::read_header_tables( |
| 248 | const unsigned char* lineptr) |
| 249 | { |
| 250 | ++this->current_header_index_; |
| 251 | |
| 252 | // Create a new directories_ entry and a new files_ entry for our new |
| 253 | // header. We initialize each with a single empty element, because |
| 254 | // dwarf indexes directory and filenames starting at 1. |
| 255 | gold_assert(static_cast<int>(this->directories_.size()) |
| 256 | == this->current_header_index_); |
| 257 | gold_assert(static_cast<int>(this->files_.size()) |
| 258 | == this->current_header_index_); |
| 259 | this->directories_.push_back(std::vector<std::string>(1)); |
| 260 | this->files_.push_back(std::vector<std::pair<int, std::string> >(1)); |
| 261 | |
| 262 | // It is legal for the directory entry table to be empty. |
| 263 | if (*lineptr) |
| 264 | { |
| 265 | int dirindex = 1; |
| 266 | while (*lineptr) |
| 267 | { |
| 268 | const char* dirname = reinterpret_cast<const char*>(lineptr); |
| 269 | gold_assert(dirindex |
| 270 | == static_cast<int>(this->directories_.back().size())); |
| 271 | this->directories_.back().push_back(dirname); |
| 272 | lineptr += this->directories_.back().back().size() + 1; |
| 273 | dirindex++; |
| 274 | } |
| 275 | } |
| 276 | lineptr++; |
| 277 | |
| 278 | // It is also legal for the file entry table to be empty. |
| 279 | if (*lineptr) |
| 280 | { |
| 281 | int fileindex = 1; |
| 282 | size_t len; |
| 283 | while (*lineptr) |
| 284 | { |
| 285 | const char* filename = reinterpret_cast<const char*>(lineptr); |
| 286 | lineptr += strlen(filename) + 1; |
| 287 | |
| 288 | uint64_t dirindex = read_unsigned_LEB_128(lineptr, &len); |
| 289 | lineptr += len; |
| 290 | |
| 291 | if (dirindex >= this->directories_.back().size()) |
| 292 | dirindex = 0; |
| 293 | int dirindexi = static_cast<int>(dirindex); |
| 294 | |
| 295 | read_unsigned_LEB_128(lineptr, &len); // mod_time |
| 296 | lineptr += len; |
| 297 | |
| 298 | read_unsigned_LEB_128(lineptr, &len); // filelength |
| 299 | lineptr += len; |
| 300 | |
| 301 | gold_assert(fileindex |
| 302 | == static_cast<int>(this->files_.back().size())); |
| 303 | this->files_.back().push_back(std::make_pair(dirindexi, filename)); |
| 304 | fileindex++; |
| 305 | } |
| 306 | } |
| 307 | lineptr++; |
| 308 | |
| 309 | return lineptr; |
| 310 | } |
| 311 | |
| 312 | // Process a single opcode in the .debug.line structure. |
| 313 | |
| 314 | // Templating on size and big_endian would yield more efficient (and |
| 315 | // simpler) code, but would bloat the binary. Speed isn't important |
| 316 | // here. |
| 317 | |
| 318 | template<int size, bool big_endian> |
| 319 | bool |
| 320 | Sized_dwarf_line_info<size, big_endian>::process_one_opcode( |
| 321 | const unsigned char* start, struct LineStateMachine* lsm, size_t* len) |
| 322 | { |
| 323 | size_t oplen = 0; |
| 324 | size_t templen; |
| 325 | unsigned char opcode = *start; |
| 326 | oplen++; |
| 327 | start++; |
| 328 | |
| 329 | // If the opcode is great than the opcode_base, it is a special |
| 330 | // opcode. Most line programs consist mainly of special opcodes. |
| 331 | if (opcode >= header_.opcode_base) |
| 332 | { |
| 333 | opcode -= header_.opcode_base; |
| 334 | const int advance_address = ((opcode / header_.line_range) |
| 335 | * header_.min_insn_length); |
| 336 | lsm->address += advance_address; |
| 337 | |
| 338 | const int advance_line = ((opcode % header_.line_range) |
| 339 | + header_.line_base); |
| 340 | lsm->line_num += advance_line; |
| 341 | lsm->basic_block = true; |
| 342 | *len = oplen; |
| 343 | return true; |
| 344 | } |
| 345 | |
| 346 | // Otherwise, we have the regular opcodes |
| 347 | switch (opcode) |
| 348 | { |
| 349 | case elfcpp::DW_LNS_copy: |
| 350 | lsm->basic_block = false; |
| 351 | *len = oplen; |
| 352 | return true; |
| 353 | |
| 354 | case elfcpp::DW_LNS_advance_pc: |
| 355 | { |
| 356 | const uint64_t advance_address |
| 357 | = read_unsigned_LEB_128(start, &templen); |
| 358 | oplen += templen; |
| 359 | lsm->address += header_.min_insn_length * advance_address; |
| 360 | } |
| 361 | break; |
| 362 | |
| 363 | case elfcpp::DW_LNS_advance_line: |
| 364 | { |
| 365 | const uint64_t advance_line = read_signed_LEB_128(start, &templen); |
| 366 | oplen += templen; |
| 367 | lsm->line_num += advance_line; |
| 368 | } |
| 369 | break; |
| 370 | |
| 371 | case elfcpp::DW_LNS_set_file: |
| 372 | { |
| 373 | const uint64_t fileno = read_unsigned_LEB_128(start, &templen); |
| 374 | oplen += templen; |
| 375 | lsm->file_num = fileno; |
| 376 | } |
| 377 | break; |
| 378 | |
| 379 | case elfcpp::DW_LNS_set_column: |
| 380 | { |
| 381 | const uint64_t colno = read_unsigned_LEB_128(start, &templen); |
| 382 | oplen += templen; |
| 383 | lsm->column_num = colno; |
| 384 | } |
| 385 | break; |
| 386 | |
| 387 | case elfcpp::DW_LNS_negate_stmt: |
| 388 | lsm->is_stmt = !lsm->is_stmt; |
| 389 | break; |
| 390 | |
| 391 | case elfcpp::DW_LNS_set_basic_block: |
| 392 | lsm->basic_block = true; |
| 393 | break; |
| 394 | |
| 395 | case elfcpp::DW_LNS_fixed_advance_pc: |
| 396 | { |
| 397 | int advance_address; |
| 398 | advance_address = elfcpp::Swap_unaligned<16, big_endian>::readval(start); |
| 399 | oplen += 2; |
| 400 | lsm->address += advance_address; |
| 401 | } |
| 402 | break; |
| 403 | |
| 404 | case elfcpp::DW_LNS_const_add_pc: |
| 405 | { |
| 406 | const int advance_address = (header_.min_insn_length |
| 407 | * ((255 - header_.opcode_base) |
| 408 | / header_.line_range)); |
| 409 | lsm->address += advance_address; |
| 410 | } |
| 411 | break; |
| 412 | |
| 413 | case elfcpp::DW_LNS_extended_op: |
| 414 | { |
| 415 | const uint64_t extended_op_len |
| 416 | = read_unsigned_LEB_128(start, &templen); |
| 417 | start += templen; |
| 418 | oplen += templen + extended_op_len; |
| 419 | |
| 420 | const unsigned char extended_op = *start; |
| 421 | start++; |
| 422 | |
| 423 | switch (extended_op) |
| 424 | { |
| 425 | case elfcpp::DW_LNE_end_sequence: |
| 426 | // This means that the current byte is the one immediately |
| 427 | // after a set of instructions. Record the current line |
| 428 | // for up to one less than the current address. |
| 429 | lsm->line_num = -1; |
| 430 | lsm->end_sequence = true; |
| 431 | *len = oplen; |
| 432 | return true; |
| 433 | |
| 434 | case elfcpp::DW_LNE_set_address: |
| 435 | { |
| 436 | lsm->address = elfcpp::Swap_unaligned<size, big_endian>::readval(start); |
| 437 | typename Reloc_map::const_iterator it |
| 438 | = reloc_map_.find(start - this->buffer_); |
| 439 | if (it != reloc_map_.end()) |
| 440 | { |
| 441 | // value + addend. |
| 442 | lsm->address += it->second.second; |
| 443 | lsm->shndx = it->second.first; |
| 444 | } |
| 445 | else |
| 446 | { |
| 447 | // If we're a normal .o file, with relocs, every |
| 448 | // set_address should have an associated relocation. |
| 449 | if (this->input_is_relobj()) |
| 450 | this->data_valid_ = false; |
| 451 | } |
| 452 | break; |
| 453 | } |
| 454 | case elfcpp::DW_LNE_define_file: |
| 455 | { |
| 456 | const char* filename = reinterpret_cast<const char*>(start); |
| 457 | templen = strlen(filename) + 1; |
| 458 | start += templen; |
| 459 | |
| 460 | uint64_t dirindex = read_unsigned_LEB_128(start, &templen); |
| 461 | oplen += templen; |
| 462 | |
| 463 | if (dirindex >= this->directories_.back().size()) |
| 464 | dirindex = 0; |
| 465 | int dirindexi = static_cast<int>(dirindex); |
| 466 | |
| 467 | read_unsigned_LEB_128(start, &templen); // mod_time |
| 468 | oplen += templen; |
| 469 | |
| 470 | read_unsigned_LEB_128(start, &templen); // filelength |
| 471 | oplen += templen; |
| 472 | |
| 473 | this->files_.back().push_back(std::make_pair(dirindexi, |
| 474 | filename)); |
| 475 | } |
| 476 | break; |
| 477 | } |
| 478 | } |
| 479 | break; |
| 480 | |
| 481 | default: |
| 482 | { |
| 483 | // Ignore unknown opcode silently |
| 484 | for (int i = 0; i < header_.std_opcode_lengths[opcode]; i++) |
| 485 | { |
| 486 | size_t templen; |
| 487 | read_unsigned_LEB_128(start, &templen); |
| 488 | start += templen; |
| 489 | oplen += templen; |
| 490 | } |
| 491 | } |
| 492 | break; |
| 493 | } |
| 494 | *len = oplen; |
| 495 | return false; |
| 496 | } |
| 497 | |
| 498 | // Read the debug information at LINEPTR and store it in the line |
| 499 | // number map. |
| 500 | |
| 501 | template<int size, bool big_endian> |
| 502 | unsigned const char* |
| 503 | Sized_dwarf_line_info<size, big_endian>::read_lines(unsigned const char* lineptr, |
| 504 | off_t shndx) |
| 505 | { |
| 506 | struct LineStateMachine lsm; |
| 507 | |
| 508 | // LENGTHSTART is the place the length field is based on. It is the |
| 509 | // point in the header after the initial length field. |
| 510 | const unsigned char* lengthstart = buffer_; |
| 511 | |
| 512 | // In 64 bit dwarf, the initial length is 12 bytes, because of the |
| 513 | // 0xffffffff at the start. |
| 514 | if (header_.offset_size == 8) |
| 515 | lengthstart += 12; |
| 516 | else |
| 517 | lengthstart += 4; |
| 518 | |
| 519 | while (lineptr < lengthstart + header_.total_length) |
| 520 | { |
| 521 | ResetLineStateMachine(&lsm, header_.default_is_stmt); |
| 522 | while (!lsm.end_sequence) |
| 523 | { |
| 524 | size_t oplength; |
| 525 | bool add_line = this->process_one_opcode(lineptr, &lsm, &oplength); |
| 526 | if (add_line |
| 527 | && (shndx == -1U || lsm.shndx == -1U || shndx == lsm.shndx)) |
| 528 | { |
| 529 | Offset_to_lineno_entry entry |
| 530 | = { lsm.address, this->current_header_index_, |
| 531 | lsm.file_num, lsm.line_num }; |
| 532 | line_number_map_[lsm.shndx].push_back(entry); |
| 533 | } |
| 534 | lineptr += oplength; |
| 535 | } |
| 536 | } |
| 537 | |
| 538 | return lengthstart + header_.total_length; |
| 539 | } |
| 540 | |
| 541 | // Looks in the symtab to see what section a symbol is in. |
| 542 | |
| 543 | template<int size, bool big_endian> |
| 544 | unsigned int |
| 545 | Sized_dwarf_line_info<size, big_endian>::symbol_section( |
| 546 | Object* object, |
| 547 | unsigned int sym, |
| 548 | typename elfcpp::Elf_types<size>::Elf_Addr* value, |
| 549 | bool* is_ordinary) |
| 550 | { |
| 551 | const int symsize = elfcpp::Elf_sizes<size>::sym_size; |
| 552 | gold_assert(sym * symsize < this->symtab_buffer_size_); |
| 553 | elfcpp::Sym<size, big_endian> elfsym(this->symtab_buffer_ + sym * symsize); |
| 554 | *value = elfsym.get_st_value(); |
| 555 | return object->adjust_sym_shndx(sym, elfsym.get_st_shndx(), is_ordinary); |
| 556 | } |
| 557 | |
| 558 | // Read the relocations into a Reloc_map. |
| 559 | |
| 560 | template<int size, bool big_endian> |
| 561 | void |
| 562 | Sized_dwarf_line_info<size, big_endian>::read_relocs(Object* object) |
| 563 | { |
| 564 | if (this->symtab_buffer_ == NULL) |
| 565 | return; |
| 566 | |
| 567 | typename elfcpp::Elf_types<size>::Elf_Addr value; |
| 568 | off_t reloc_offset; |
| 569 | while ((reloc_offset = this->track_relocs_.next_offset()) != -1) |
| 570 | { |
| 571 | const unsigned int sym = this->track_relocs_.next_symndx(); |
| 572 | |
| 573 | bool is_ordinary; |
| 574 | const unsigned int shndx = this->symbol_section(object, sym, &value, |
| 575 | &is_ordinary); |
| 576 | |
| 577 | // There is no reason to record non-ordinary section indexes, or |
| 578 | // SHN_UNDEF, because they will never match the real section. |
| 579 | if (is_ordinary && shndx != elfcpp::SHN_UNDEF) |
| 580 | this->reloc_map_[reloc_offset] = std::make_pair(shndx, value); |
| 581 | |
| 582 | this->track_relocs_.advance(reloc_offset + 1); |
| 583 | } |
| 584 | } |
| 585 | |
| 586 | // Read the line number info. |
| 587 | |
| 588 | template<int size, bool big_endian> |
| 589 | void |
| 590 | Sized_dwarf_line_info<size, big_endian>::read_line_mappings(Object* object, |
| 591 | off_t shndx) |
| 592 | { |
| 593 | gold_assert(this->data_valid_ == true); |
| 594 | |
| 595 | this->read_relocs(object); |
| 596 | while (this->buffer_ < this->buffer_end_) |
| 597 | { |
| 598 | const unsigned char* lineptr = this->buffer_; |
| 599 | lineptr = this->read_header_prolog(lineptr); |
| 600 | lineptr = this->read_header_tables(lineptr); |
| 601 | lineptr = this->read_lines(lineptr, shndx); |
| 602 | this->buffer_ = lineptr; |
| 603 | } |
| 604 | |
| 605 | // Sort the lines numbers, so addr2line can use binary search. |
| 606 | for (typename Lineno_map::iterator it = line_number_map_.begin(); |
| 607 | it != line_number_map_.end(); |
| 608 | ++it) |
| 609 | // Each vector needs to be sorted by offset. |
| 610 | std::sort(it->second.begin(), it->second.end()); |
| 611 | } |
| 612 | |
| 613 | // Some processing depends on whether the input is a .o file or not. |
| 614 | // For instance, .o files have relocs, and have .debug_lines |
| 615 | // information on a per section basis. .so files, on the other hand, |
| 616 | // lack relocs, and offsets are unique, so we can ignore the section |
| 617 | // information. |
| 618 | |
| 619 | template<int size, bool big_endian> |
| 620 | bool |
| 621 | Sized_dwarf_line_info<size, big_endian>::input_is_relobj() |
| 622 | { |
| 623 | // Only .o files have relocs and the symtab buffer that goes with them. |
| 624 | return this->symtab_buffer_ != NULL; |
| 625 | } |
| 626 | |
| 627 | // Given an Offset_to_lineno_entry vector, and an offset, figure out |
| 628 | // if the offset points into a function according to the vector (see |
| 629 | // comments below for the algorithm). If it does, return an iterator |
| 630 | // into the vector that points to the line-number that contains that |
| 631 | // offset. If not, it returns vector::end(). |
| 632 | |
| 633 | static std::vector<Offset_to_lineno_entry>::const_iterator |
| 634 | offset_to_iterator(const std::vector<Offset_to_lineno_entry>* offsets, |
| 635 | off_t offset) |
| 636 | { |
| 637 | const Offset_to_lineno_entry lookup_key = { offset, 0, 0, 0 }; |
| 638 | |
| 639 | // lower_bound() returns the smallest offset which is >= lookup_key. |
| 640 | // If no offset in offsets is >= lookup_key, returns end(). |
| 641 | std::vector<Offset_to_lineno_entry>::const_iterator it |
| 642 | = std::lower_bound(offsets->begin(), offsets->end(), lookup_key); |
| 643 | |
| 644 | // This code is easiest to understand with a concrete example. |
| 645 | // Here's a possible offsets array: |
| 646 | // {{offset = 3211, header_num = 0, file_num = 1, line_num = 16}, // 0 |
| 647 | // {offset = 3224, header_num = 0, file_num = 1, line_num = 20}, // 1 |
| 648 | // {offset = 3226, header_num = 0, file_num = 1, line_num = 22}, // 2 |
| 649 | // {offset = 3231, header_num = 0, file_num = 1, line_num = 25}, // 3 |
| 650 | // {offset = 3232, header_num = 0, file_num = 1, line_num = -1}, // 4 |
| 651 | // {offset = 3232, header_num = 0, file_num = 1, line_num = 65}, // 5 |
| 652 | // {offset = 3235, header_num = 0, file_num = 1, line_num = 66}, // 6 |
| 653 | // {offset = 3236, header_num = 0, file_num = 1, line_num = -1}, // 7 |
| 654 | // {offset = 5764, header_num = 0, file_num = 1, line_num = 47}, // 8 |
| 655 | // {offset = 5765, header_num = 0, file_num = 1, line_num = 48}, // 9 |
| 656 | // {offset = 5767, header_num = 0, file_num = 1, line_num = 49}, // 10 |
| 657 | // {offset = 5768, header_num = 0, file_num = 1, line_num = 50}, // 11 |
| 658 | // {offset = 5773, header_num = 0, file_num = 1, line_num = -1}, // 12 |
| 659 | // {offset = 5787, header_num = 1, file_num = 1, line_num = 19}, // 13 |
| 660 | // {offset = 5790, header_num = 1, file_num = 1, line_num = 20}, // 14 |
| 661 | // {offset = 5793, header_num = 1, file_num = 1, line_num = 67}, // 15 |
| 662 | // {offset = 5793, header_num = 1, file_num = 1, line_num = -1}, // 16 |
| 663 | // {offset = 5795, header_num = 1, file_num = 1, line_num = 68}, // 17 |
| 664 | // {offset = 5798, header_num = 1, file_num = 1, line_num = -1}, // 18 |
| 665 | // The entries with line_num == -1 mark the end of a function: the |
| 666 | // associated offset is one past the last instruction in the |
| 667 | // function. This can correspond to the beginning of the next |
| 668 | // function (as is true for offset 3232); alternately, there can be |
| 669 | // a gap between the end of one function and the start of the next |
| 670 | // (as is true for some others, most obviously from 3236->5764). |
| 671 | // |
| 672 | // Case 1: lookup_key has offset == 10. lower_bound returns |
| 673 | // offsets[0]. Since it's not an exact match and we're |
| 674 | // at the beginning of offsets, we return end() (invalid). |
| 675 | // Case 2: lookup_key has offset 10000. lower_bound returns |
| 676 | // offset[19] (end()). We return end() (invalid). |
| 677 | // Case 3: lookup_key has offset == 3211. lower_bound matches |
| 678 | // offsets[0] exactly, and that's the entry we return. |
| 679 | // Case 4: lookup_key has offset == 3232. lower_bound returns |
| 680 | // offsets[4]. That's an exact match, but indicates |
| 681 | // end-of-function. We check if offsets[5] is also an |
| 682 | // exact match but not end-of-function. It is, so we |
| 683 | // return offsets[5]. |
| 684 | // Case 5: lookup_key has offset == 3214. lower_bound returns |
| 685 | // offsets[1]. Since it's not an exact match, we back |
| 686 | // up to the offset that's < lookup_key, offsets[0]. |
| 687 | // We note offsets[0] is a valid entry (not end-of-function), |
| 688 | // so that's the entry we return. |
| 689 | // Case 6: lookup_key has offset == 4000. lower_bound returns |
| 690 | // offsets[8]. Since it's not an exact match, we back |
| 691 | // up to offsets[7]. Since offsets[7] indicates |
| 692 | // end-of-function, we know lookup_key is between |
| 693 | // functions, so we return end() (not a valid offset). |
| 694 | // Case 7: lookup_key has offset == 5794. lower_bound returns |
| 695 | // offsets[17]. Since it's not an exact match, we back |
| 696 | // up to offsets[15]. Note we back up to the *first* |
| 697 | // entry with offset 5793, not just offsets[17-1]. |
| 698 | // We note offsets[15] is a valid entry, so we return it. |
| 699 | // If offsets[15] had had line_num == -1, we would have |
| 700 | // checked offsets[16]. The reason for this is that |
| 701 | // 15 and 16 can be in an arbitrary order, since we sort |
| 702 | // only by offset. (Note it doesn't help to use line_number |
| 703 | // as a secondary sort key, since sometimes we want the -1 |
| 704 | // to be first and sometimes we want it to be last.) |
| 705 | |
| 706 | // This deals with cases (1) and (2). |
| 707 | if ((it == offsets->begin() && offset < it->offset) |
| 708 | || it == offsets->end()) |
| 709 | return offsets->end(); |
| 710 | |
| 711 | // This deals with cases (3) and (4). |
| 712 | if (offset == it->offset) |
| 713 | { |
| 714 | while (it != offsets->end() |
| 715 | && it->offset == offset |
| 716 | && it->line_num == -1) |
| 717 | ++it; |
| 718 | if (it == offsets->end() || it->offset != offset) |
| 719 | return offsets->end(); |
| 720 | else |
| 721 | return it; |
| 722 | } |
| 723 | |
| 724 | // This handles the first part of case (7) -- we back up to the |
| 725 | // *first* entry that has the offset that's behind us. |
| 726 | gold_assert(it != offsets->begin()); |
| 727 | std::vector<Offset_to_lineno_entry>::const_iterator range_end = it; |
| 728 | --it; |
| 729 | const off_t range_value = it->offset; |
| 730 | while (it != offsets->begin() && (it-1)->offset == range_value) |
| 731 | --it; |
| 732 | |
| 733 | // This handles cases (5), (6), and (7): if any entry in the |
| 734 | // equal_range [it, range_end) has a line_num != -1, it's a valid |
| 735 | // match. If not, we're not in a function. |
| 736 | for (; it != range_end; ++it) |
| 737 | if (it->line_num != -1) |
| 738 | return it; |
| 739 | return offsets->end(); |
| 740 | } |
| 741 | |
| 742 | // Return a string for a file name and line number. |
| 743 | |
| 744 | template<int size, bool big_endian> |
| 745 | std::string |
| 746 | Sized_dwarf_line_info<size, big_endian>::do_addr2line(unsigned int shndx, |
| 747 | off_t offset) |
| 748 | { |
| 749 | if (this->data_valid_ == false) |
| 750 | return ""; |
| 751 | |
| 752 | const std::vector<Offset_to_lineno_entry>* offsets; |
| 753 | // If we do not have reloc information, then our input is a .so or |
| 754 | // some similar data structure where all the information is held in |
| 755 | // the offset. In that case, we ignore the input shndx. |
| 756 | if (this->input_is_relobj()) |
| 757 | offsets = &this->line_number_map_[shndx]; |
| 758 | else |
| 759 | offsets = &this->line_number_map_[-1U]; |
| 760 | if (offsets->empty()) |
| 761 | return ""; |
| 762 | |
| 763 | typename std::vector<Offset_to_lineno_entry>::const_iterator it |
| 764 | = offset_to_iterator(offsets, offset); |
| 765 | if (it == offsets->end()) |
| 766 | return ""; |
| 767 | |
| 768 | // Convert the file_num + line_num into a string. |
| 769 | std::string ret; |
| 770 | |
| 771 | gold_assert(it->header_num < static_cast<int>(this->files_.size())); |
| 772 | gold_assert(it->file_num |
| 773 | < static_cast<int>(this->files_[it->header_num].size())); |
| 774 | const std::pair<int, std::string>& filename_pair |
| 775 | = this->files_[it->header_num][it->file_num]; |
| 776 | const std::string& filename = filename_pair.second; |
| 777 | |
| 778 | gold_assert(it->header_num < static_cast<int>(this->directories_.size())); |
| 779 | gold_assert(filename_pair.first |
| 780 | < static_cast<int>(this->directories_[it->header_num].size())); |
| 781 | const std::string& dirname |
| 782 | = this->directories_[it->header_num][filename_pair.first]; |
| 783 | |
| 784 | if (!dirname.empty()) |
| 785 | { |
| 786 | ret += dirname; |
| 787 | ret += "/"; |
| 788 | } |
| 789 | ret += filename; |
| 790 | if (ret.empty()) |
| 791 | ret = "(unknown)"; |
| 792 | |
| 793 | char buffer[64]; // enough to hold a line number |
| 794 | snprintf(buffer, sizeof(buffer), "%d", it->line_num); |
| 795 | ret += ":"; |
| 796 | ret += buffer; |
| 797 | |
| 798 | return ret; |
| 799 | } |
| 800 | |
| 801 | // Dwarf_line_info routines. |
| 802 | |
| 803 | static unsigned int next_generation_count = 0; |
| 804 | |
| 805 | struct Addr2line_cache_entry |
| 806 | { |
| 807 | Object* object; |
| 808 | unsigned int shndx; |
| 809 | Dwarf_line_info* dwarf_line_info; |
| 810 | unsigned int generation_count; |
| 811 | unsigned int access_count; |
| 812 | |
| 813 | Addr2line_cache_entry(Object* o, unsigned int s, Dwarf_line_info* d) |
| 814 | : object(o), shndx(s), dwarf_line_info(d), |
| 815 | generation_count(next_generation_count), access_count(0) |
| 816 | { |
| 817 | if (next_generation_count < (1U << 31)) |
| 818 | ++next_generation_count; |
| 819 | } |
| 820 | }; |
| 821 | // We expect this cache to be small, so don't bother with a hashtable |
| 822 | // or priority queue or anything: just use a simple vector. |
| 823 | static std::vector<Addr2line_cache_entry> addr2line_cache; |
| 824 | |
| 825 | std::string |
| 826 | Dwarf_line_info::one_addr2line(Object* object, |
| 827 | unsigned int shndx, off_t offset, |
| 828 | size_t cache_size) |
| 829 | { |
| 830 | Dwarf_line_info* lineinfo = NULL; |
| 831 | std::vector<Addr2line_cache_entry>::iterator it; |
| 832 | |
| 833 | // First, check the cache. If we hit, update the counts. |
| 834 | for (it = addr2line_cache.begin(); it != addr2line_cache.end(); ++it) |
| 835 | { |
| 836 | if (it->object == object && it->shndx == shndx) |
| 837 | { |
| 838 | lineinfo = it->dwarf_line_info; |
| 839 | it->generation_count = next_generation_count; |
| 840 | // We cap generation_count at 2^31 -1 to avoid overflow. |
| 841 | if (next_generation_count < (1U << 31)) |
| 842 | ++next_generation_count; |
| 843 | // We cap access_count at 31 so 2^access_count doesn't overflow |
| 844 | if (it->access_count < 31) |
| 845 | ++it->access_count; |
| 846 | break; |
| 847 | } |
| 848 | } |
| 849 | |
| 850 | // If we don't hit the cache, create a new object and insert into the |
| 851 | // cache. |
| 852 | if (lineinfo == NULL) |
| 853 | { |
| 854 | switch (parameters->size_and_endianness()) |
| 855 | { |
| 856 | #ifdef HAVE_TARGET_32_LITTLE |
| 857 | case Parameters::TARGET_32_LITTLE: |
| 858 | lineinfo = new Sized_dwarf_line_info<32, false>(object, shndx); break; |
| 859 | #endif |
| 860 | #ifdef HAVE_TARGET_32_BIG |
| 861 | case Parameters::TARGET_32_BIG: |
| 862 | lineinfo = new Sized_dwarf_line_info<32, true>(object, shndx); break; |
| 863 | #endif |
| 864 | #ifdef HAVE_TARGET_64_LITTLE |
| 865 | case Parameters::TARGET_64_LITTLE: |
| 866 | lineinfo = new Sized_dwarf_line_info<64, false>(object, shndx); break; |
| 867 | #endif |
| 868 | #ifdef HAVE_TARGET_64_BIG |
| 869 | case Parameters::TARGET_64_BIG: |
| 870 | lineinfo = new Sized_dwarf_line_info<64, true>(object, shndx); break; |
| 871 | #endif |
| 872 | default: |
| 873 | gold_unreachable(); |
| 874 | } |
| 875 | addr2line_cache.push_back(Addr2line_cache_entry(object, shndx, lineinfo)); |
| 876 | } |
| 877 | |
| 878 | // Now that we have our object, figure out the answer |
| 879 | std::string retval = lineinfo->addr2line(shndx, offset); |
| 880 | |
| 881 | // Finally, if our cache has grown too big, delete old objects. We |
| 882 | // assume the common (probably only) case is deleting only one object. |
| 883 | // We use a pretty simple scheme to evict: function of LRU and MFU. |
| 884 | while (addr2line_cache.size() > cache_size) |
| 885 | { |
| 886 | unsigned int lowest_score = ~0U; |
| 887 | std::vector<Addr2line_cache_entry>::iterator lowest |
| 888 | = addr2line_cache.end(); |
| 889 | for (it = addr2line_cache.begin(); it != addr2line_cache.end(); ++it) |
| 890 | { |
| 891 | const unsigned int score = (it->generation_count |
| 892 | + (1U << it->access_count)); |
| 893 | if (score < lowest_score) |
| 894 | { |
| 895 | lowest_score = score; |
| 896 | lowest = it; |
| 897 | } |
| 898 | } |
| 899 | if (lowest != addr2line_cache.end()) |
| 900 | { |
| 901 | delete lowest->dwarf_line_info; |
| 902 | addr2line_cache.erase(lowest); |
| 903 | } |
| 904 | } |
| 905 | |
| 906 | return retval; |
| 907 | } |
| 908 | |
| 909 | void |
| 910 | Dwarf_line_info::clear_addr2line_cache() |
| 911 | { |
| 912 | for (std::vector<Addr2line_cache_entry>::iterator it = addr2line_cache.begin(); |
| 913 | it != addr2line_cache.end(); |
| 914 | ++it) |
| 915 | delete it->dwarf_line_info; |
| 916 | addr2line_cache.clear(); |
| 917 | } |
| 918 | |
| 919 | #ifdef HAVE_TARGET_32_LITTLE |
| 920 | template |
| 921 | class Sized_dwarf_line_info<32, false>; |
| 922 | #endif |
| 923 | |
| 924 | #ifdef HAVE_TARGET_32_BIG |
| 925 | template |
| 926 | class Sized_dwarf_line_info<32, true>; |
| 927 | #endif |
| 928 | |
| 929 | #ifdef HAVE_TARGET_64_LITTLE |
| 930 | template |
| 931 | class Sized_dwarf_line_info<64, false>; |
| 932 | #endif |
| 933 | |
| 934 | #ifdef HAVE_TARGET_64_BIG |
| 935 | template |
| 936 | class Sized_dwarf_line_info<64, true>; |
| 937 | #endif |
| 938 | |
| 939 | } // End namespace gold. |