1 // dwarf_reader.cc -- parse dwarf2/3 debug information
3 // Copyright 2007, 2008 Free Software Foundation, Inc.
4 // Written by Ian Lance Taylor <iant@google.com>.
6 // This file is part of gold.
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.
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.
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.
27 #include "elfcpp_swap.h"
30 #include "parameters.h"
32 #include "dwarf_reader.h"
36 // Read an unsigned LEB128 number. Each byte contains 7 bits of
37 // information, plus one bit saying whether the number continues or
41 read_unsigned_LEB_128(const unsigned char* buffer, size_t* len)
45 unsigned int shift = 0;
52 result |= (static_cast<uint64_t>(byte & 0x7f)) << shift;
62 // Read a signed LEB128 number. These are like regular LEB128
63 // numbers, except the last byte may have a sign bit set.
66 read_signed_LEB_128(const unsigned char* buffer, size_t* len)
77 result |= (static_cast<uint64_t>(byte & 0x7f) << shift);
82 if ((shift < 8 * static_cast<int>(sizeof(result))) && (byte & 0x40))
83 result |= -((static_cast<int64_t>(1)) << shift);
88 } // End anonymous namespace.
93 // This is the format of a DWARF2/3 line state machine that we process
94 // opcodes using. There is no need for anything outside the lineinfo
95 // processor to know how this works.
97 struct LineStateMachine
103 unsigned int shndx; // the section address refers to
104 bool is_stmt; // stmt means statement.
110 ResetLineStateMachine(struct LineStateMachine* lsm, bool default_is_stmt)
117 lsm->is_stmt = default_is_stmt;
118 lsm->basic_block = false;
119 lsm->end_sequence = false;
122 template<int size, bool big_endian>
123 Sized_dwarf_line_info<size, big_endian>::Sized_dwarf_line_info(Object* object,
125 : data_valid_(false), buffer_(NULL), symtab_buffer_(NULL),
126 directories_(), files_(), current_header_index_(-1)
128 unsigned int debug_shndx;
129 for (debug_shndx = 0; debug_shndx < object->shnum(); ++debug_shndx)
130 // FIXME: do this more efficiently: section_name() isn't super-fast
131 if (object->section_name(debug_shndx) == ".debug_line")
133 section_size_type buffer_size;
134 this->buffer_ = object->section_contents(debug_shndx, &buffer_size,
136 this->buffer_end_ = this->buffer_ + buffer_size;
139 if (this->buffer_ == NULL)
142 // Find the relocation section for ".debug_line".
143 // We expect these for relobjs (.o's) but not dynobjs (.so's).
144 bool got_relocs = false;
145 for (unsigned int reloc_shndx = 0;
146 reloc_shndx < object->shnum();
149 unsigned int reloc_sh_type = object->section_type(reloc_shndx);
150 if ((reloc_sh_type == elfcpp::SHT_REL
151 || reloc_sh_type == elfcpp::SHT_RELA)
152 && object->section_info(reloc_shndx) == debug_shndx)
154 got_relocs = this->track_relocs_.initialize(object, reloc_shndx,
160 // Finally, we need the symtab section to interpret the relocs.
163 unsigned int symtab_shndx;
164 for (symtab_shndx = 0; symtab_shndx < object->shnum(); ++symtab_shndx)
165 if (object->section_type(symtab_shndx) == elfcpp::SHT_SYMTAB)
167 this->symtab_buffer_ = object->section_contents(
168 symtab_shndx, &this->symtab_buffer_size_, false);
171 if (this->symtab_buffer_ == NULL)
175 // Now that we have successfully read all the data, parse the debug
177 this->data_valid_ = true;
178 this->read_line_mappings(object, read_shndx);
181 // Read the DWARF header.
183 template<int size, bool big_endian>
185 Sized_dwarf_line_info<size, big_endian>::read_header_prolog(
186 const unsigned char* lineptr)
188 uint32_t initial_length = elfcpp::Swap_unaligned<32, big_endian>::readval(lineptr);
191 // In DWARF2/3, if the initial length is all 1 bits, then the offset
192 // size is 8 and we need to read the next 8 bytes for the real length.
193 if (initial_length == 0xffffffff)
195 header_.offset_size = 8;
196 initial_length = elfcpp::Swap_unaligned<64, big_endian>::readval(lineptr);
200 header_.offset_size = 4;
202 header_.total_length = initial_length;
204 gold_assert(lineptr + header_.total_length <= buffer_end_);
206 header_.version = elfcpp::Swap_unaligned<16, big_endian>::readval(lineptr);
209 if (header_.offset_size == 4)
210 header_.prologue_length = elfcpp::Swap_unaligned<32, big_endian>::readval(lineptr);
212 header_.prologue_length = elfcpp::Swap_unaligned<64, big_endian>::readval(lineptr);
213 lineptr += header_.offset_size;
215 header_.min_insn_length = *lineptr;
218 header_.default_is_stmt = *lineptr;
221 header_.line_base = *reinterpret_cast<const signed char*>(lineptr);
224 header_.line_range = *lineptr;
227 header_.opcode_base = *lineptr;
230 header_.std_opcode_lengths.reserve(header_.opcode_base + 1);
231 header_.std_opcode_lengths[0] = 0;
232 for (int i = 1; i < header_.opcode_base; i++)
234 header_.std_opcode_lengths[i] = *lineptr;
241 // The header for a debug_line section is mildly complicated, because
242 // the line info is very tightly encoded.
244 template<int size, bool big_endian>
246 Sized_dwarf_line_info<size, big_endian>::read_header_tables(
247 const unsigned char* lineptr)
249 ++this->current_header_index_;
251 // Create a new directories_ entry and a new files_ entry for our new
252 // header. We initialize each with a single empty element, because
253 // dwarf indexes directory and filenames starting at 1.
254 gold_assert(static_cast<int>(this->directories_.size())
255 == this->current_header_index_);
256 gold_assert(static_cast<int>(this->files_.size())
257 == this->current_header_index_);
258 this->directories_.push_back(std::vector<std::string>(1));
259 this->files_.push_back(std::vector<std::pair<int, std::string> >(1));
261 // It is legal for the directory entry table to be empty.
267 const char* dirname = reinterpret_cast<const char*>(lineptr);
269 == static_cast<int>(this->directories_.back().size()));
270 this->directories_.back().push_back(dirname);
271 lineptr += this->directories_.back().back().size() + 1;
277 // It is also legal for the file entry table to be empty.
284 const char* filename = reinterpret_cast<const char*>(lineptr);
285 lineptr += strlen(filename) + 1;
287 uint64_t dirindex = read_unsigned_LEB_128(lineptr, &len);
290 if (dirindex >= this->directories_.back().size())
292 int dirindexi = static_cast<int>(dirindex);
294 read_unsigned_LEB_128(lineptr, &len); // mod_time
297 read_unsigned_LEB_128(lineptr, &len); // filelength
300 gold_assert(fileindex
301 == static_cast<int>(this->files_.back().size()));
302 this->files_.back().push_back(std::make_pair(dirindexi, filename));
311 // Process a single opcode in the .debug.line structure.
313 // Templating on size and big_endian would yield more efficient (and
314 // simpler) code, but would bloat the binary. Speed isn't important
317 template<int size, bool big_endian>
319 Sized_dwarf_line_info<size, big_endian>::process_one_opcode(
320 const unsigned char* start, struct LineStateMachine* lsm, size_t* len)
324 unsigned char opcode = *start;
328 // If the opcode is great than the opcode_base, it is a special
329 // opcode. Most line programs consist mainly of special opcodes.
330 if (opcode >= header_.opcode_base)
332 opcode -= header_.opcode_base;
333 const int advance_address = ((opcode / header_.line_range)
334 * header_.min_insn_length);
335 lsm->address += advance_address;
337 const int advance_line = ((opcode % header_.line_range)
338 + header_.line_base);
339 lsm->line_num += advance_line;
340 lsm->basic_block = true;
345 // Otherwise, we have the regular opcodes
348 case elfcpp::DW_LNS_copy:
349 lsm->basic_block = false;
353 case elfcpp::DW_LNS_advance_pc:
355 const uint64_t advance_address
356 = read_unsigned_LEB_128(start, &templen);
358 lsm->address += header_.min_insn_length * advance_address;
362 case elfcpp::DW_LNS_advance_line:
364 const uint64_t advance_line = read_signed_LEB_128(start, &templen);
366 lsm->line_num += advance_line;
370 case elfcpp::DW_LNS_set_file:
372 const uint64_t fileno = read_unsigned_LEB_128(start, &templen);
374 lsm->file_num = fileno;
378 case elfcpp::DW_LNS_set_column:
380 const uint64_t colno = read_unsigned_LEB_128(start, &templen);
382 lsm->column_num = colno;
386 case elfcpp::DW_LNS_negate_stmt:
387 lsm->is_stmt = !lsm->is_stmt;
390 case elfcpp::DW_LNS_set_basic_block:
391 lsm->basic_block = true;
394 case elfcpp::DW_LNS_fixed_advance_pc:
397 advance_address = elfcpp::Swap_unaligned<16, big_endian>::readval(start);
399 lsm->address += advance_address;
403 case elfcpp::DW_LNS_const_add_pc:
405 const int advance_address = (header_.min_insn_length
406 * ((255 - header_.opcode_base)
407 / header_.line_range));
408 lsm->address += advance_address;
412 case elfcpp::DW_LNS_extended_op:
414 const uint64_t extended_op_len
415 = read_unsigned_LEB_128(start, &templen);
417 oplen += templen + extended_op_len;
419 const unsigned char extended_op = *start;
424 case elfcpp::DW_LNE_end_sequence:
425 // This means that the current byte is the one immediately
426 // after a set of instructions. Record the current line
427 // for up to one less than the current address.
429 lsm->end_sequence = true;
433 case elfcpp::DW_LNE_set_address:
435 lsm->address = elfcpp::Swap_unaligned<size, big_endian>::readval(start);
436 typename Reloc_map::const_iterator it
437 = reloc_map_.find(start - this->buffer_);
438 if (it != reloc_map_.end())
441 lsm->address += it->second.second;
442 lsm->shndx = it->second.first;
446 // If we're a normal .o file, with relocs, every
447 // set_address should have an associated relocation.
448 if (this->input_is_relobj())
449 this->data_valid_ = false;
453 case elfcpp::DW_LNE_define_file:
455 const char* filename = reinterpret_cast<const char*>(start);
456 templen = strlen(filename) + 1;
459 uint64_t dirindex = read_unsigned_LEB_128(start, &templen);
462 if (dirindex >= this->directories_.back().size())
464 int dirindexi = static_cast<int>(dirindex);
466 read_unsigned_LEB_128(start, &templen); // mod_time
469 read_unsigned_LEB_128(start, &templen); // filelength
472 this->files_.back().push_back(std::make_pair(dirindexi,
482 // Ignore unknown opcode silently
483 for (int i = 0; i < header_.std_opcode_lengths[opcode]; i++)
486 read_unsigned_LEB_128(start, &templen);
497 // Read the debug information at LINEPTR and store it in the line
500 template<int size, bool big_endian>
502 Sized_dwarf_line_info<size, big_endian>::read_lines(unsigned const char* lineptr,
505 struct LineStateMachine lsm;
507 // LENGTHSTART is the place the length field is based on. It is the
508 // point in the header after the initial length field.
509 const unsigned char* lengthstart = buffer_;
511 // In 64 bit dwarf, the initial length is 12 bytes, because of the
512 // 0xffffffff at the start.
513 if (header_.offset_size == 8)
518 while (lineptr < lengthstart + header_.total_length)
520 ResetLineStateMachine(&lsm, header_.default_is_stmt);
521 while (!lsm.end_sequence)
524 bool add_line = this->process_one_opcode(lineptr, &lsm, &oplength);
526 && (shndx == -1U || lsm.shndx == -1U || shndx == lsm.shndx))
528 Offset_to_lineno_entry entry
529 = { lsm.address, this->current_header_index_,
530 lsm.file_num, lsm.line_num };
531 line_number_map_[lsm.shndx].push_back(entry);
537 return lengthstart + header_.total_length;
540 // Looks in the symtab to see what section a symbol is in.
542 template<int size, bool big_endian>
544 Sized_dwarf_line_info<size, big_endian>::symbol_section(
547 typename elfcpp::Elf_types<size>::Elf_Addr* value,
550 const int symsize = elfcpp::Elf_sizes<size>::sym_size;
551 gold_assert(sym * symsize < this->symtab_buffer_size_);
552 elfcpp::Sym<size, big_endian> elfsym(this->symtab_buffer_ + sym * symsize);
553 *value = elfsym.get_st_value();
554 return object->adjust_sym_shndx(sym, elfsym.get_st_shndx(), is_ordinary);
557 // Read the relocations into a Reloc_map.
559 template<int size, bool big_endian>
561 Sized_dwarf_line_info<size, big_endian>::read_relocs(Object* object)
563 if (this->symtab_buffer_ == NULL)
566 typename elfcpp::Elf_types<size>::Elf_Addr value;
568 while ((reloc_offset = this->track_relocs_.next_offset()) != -1)
570 const unsigned int sym = this->track_relocs_.next_symndx();
573 const unsigned int shndx = this->symbol_section(object, sym, &value,
576 // There is no reason to record non-ordinary section indexes, or
577 // SHN_UNDEF, because they will never match the real section.
578 if (is_ordinary && shndx != elfcpp::SHN_UNDEF)
579 this->reloc_map_[reloc_offset] = std::make_pair(shndx, value);
581 this->track_relocs_.advance(reloc_offset + 1);
585 // Read the line number info.
587 template<int size, bool big_endian>
589 Sized_dwarf_line_info<size, big_endian>::read_line_mappings(Object* object,
592 gold_assert(this->data_valid_ == true);
594 this->read_relocs(object);
595 while (this->buffer_ < this->buffer_end_)
597 const unsigned char* lineptr = this->buffer_;
598 lineptr = this->read_header_prolog(lineptr);
599 lineptr = this->read_header_tables(lineptr);
600 lineptr = this->read_lines(lineptr, shndx);
601 this->buffer_ = lineptr;
604 // Sort the lines numbers, so addr2line can use binary search.
605 for (typename Lineno_map::iterator it = line_number_map_.begin();
606 it != line_number_map_.end();
608 // Each vector needs to be sorted by offset.
609 std::sort(it->second.begin(), it->second.end());
612 // Some processing depends on whether the input is a .o file or not.
613 // For instance, .o files have relocs, and have .debug_lines
614 // information on a per section basis. .so files, on the other hand,
615 // lack relocs, and offsets are unique, so we can ignore the section
618 template<int size, bool big_endian>
620 Sized_dwarf_line_info<size, big_endian>::input_is_relobj()
622 // Only .o files have relocs and the symtab buffer that goes with them.
623 return this->symtab_buffer_ != NULL;
626 // Given an Offset_to_lineno_entry vector, and an offset, figure out
627 // if the offset points into a function according to the vector (see
628 // comments below for the algorithm). If it does, return an iterator
629 // into the vector that points to the line-number that contains that
630 // offset. If not, it returns vector::end().
632 static std::vector<Offset_to_lineno_entry>::const_iterator
633 offset_to_iterator(const std::vector<Offset_to_lineno_entry>* offsets,
636 const Offset_to_lineno_entry lookup_key = { offset, 0, 0, 0 };
638 // lower_bound() returns the smallest offset which is >= lookup_key.
639 // If no offset in offsets is >= lookup_key, returns end().
640 std::vector<Offset_to_lineno_entry>::const_iterator it
641 = std::lower_bound(offsets->begin(), offsets->end(), lookup_key);
643 // This code is easiest to understand with a concrete example.
644 // Here's a possible offsets array:
645 // {{offset = 3211, header_num = 0, file_num = 1, line_num = 16}, // 0
646 // {offset = 3224, header_num = 0, file_num = 1, line_num = 20}, // 1
647 // {offset = 3226, header_num = 0, file_num = 1, line_num = 22}, // 2
648 // {offset = 3231, header_num = 0, file_num = 1, line_num = 25}, // 3
649 // {offset = 3232, header_num = 0, file_num = 1, line_num = -1}, // 4
650 // {offset = 3232, header_num = 0, file_num = 1, line_num = 65}, // 5
651 // {offset = 3235, header_num = 0, file_num = 1, line_num = 66}, // 6
652 // {offset = 3236, header_num = 0, file_num = 1, line_num = -1}, // 7
653 // {offset = 5764, header_num = 0, file_num = 1, line_num = 47}, // 8
654 // {offset = 5765, header_num = 0, file_num = 1, line_num = 48}, // 9
655 // {offset = 5767, header_num = 0, file_num = 1, line_num = 49}, // 10
656 // {offset = 5768, header_num = 0, file_num = 1, line_num = 50}, // 11
657 // {offset = 5773, header_num = 0, file_num = 1, line_num = -1}, // 12
658 // {offset = 5787, header_num = 1, file_num = 1, line_num = 19}, // 13
659 // {offset = 5790, header_num = 1, file_num = 1, line_num = 20}, // 14
660 // {offset = 5793, header_num = 1, file_num = 1, line_num = 67}, // 15
661 // {offset = 5793, header_num = 1, file_num = 1, line_num = -1}, // 16
662 // {offset = 5795, header_num = 1, file_num = 1, line_num = 68}, // 17
663 // {offset = 5798, header_num = 1, file_num = 1, line_num = -1}, // 18
664 // The entries with line_num == -1 mark the end of a function: the
665 // associated offset is one past the last instruction in the
666 // function. This can correspond to the beginning of the next
667 // function (as is true for offset 3232); alternately, there can be
668 // a gap between the end of one function and the start of the next
669 // (as is true for some others, most obviously from 3236->5764).
671 // Case 1: lookup_key has offset == 10. lower_bound returns
672 // offsets[0]. Since it's not an exact match and we're
673 // at the beginning of offsets, we return end() (invalid).
674 // Case 2: lookup_key has offset 10000. lower_bound returns
675 // offset[19] (end()). We return end() (invalid).
676 // Case 3: lookup_key has offset == 3211. lower_bound matches
677 // offsets[0] exactly, and that's the entry we return.
678 // Case 4: lookup_key has offset == 3232. lower_bound returns
679 // offsets[4]. That's an exact match, but indicates
680 // end-of-function. We check if offsets[5] is also an
681 // exact match but not end-of-function. It is, so we
682 // return offsets[5].
683 // Case 5: lookup_key has offset == 3214. lower_bound returns
684 // offsets[1]. Since it's not an exact match, we back
685 // up to the offset that's < lookup_key, offsets[0].
686 // We note offsets[0] is a valid entry (not end-of-function),
687 // so that's the entry we return.
688 // Case 6: lookup_key has offset == 4000. lower_bound returns
689 // offsets[8]. Since it's not an exact match, we back
690 // up to offsets[7]. Since offsets[7] indicates
691 // end-of-function, we know lookup_key is between
692 // functions, so we return end() (not a valid offset).
693 // Case 7: lookup_key has offset == 5794. lower_bound returns
694 // offsets[17]. Since it's not an exact match, we back
695 // up to offsets[15]. Note we back up to the *first*
696 // entry with offset 5793, not just offsets[17-1].
697 // We note offsets[15] is a valid entry, so we return it.
698 // If offsets[15] had had line_num == -1, we would have
699 // checked offsets[16]. The reason for this is that
700 // 15 and 16 can be in an arbitrary order, since we sort
701 // only by offset. (Note it doesn't help to use line_number
702 // as a secondary sort key, since sometimes we want the -1
703 // to be first and sometimes we want it to be last.)
705 // This deals with cases (1) and (2).
706 if ((it == offsets->begin() && offset < it->offset)
707 || it == offsets->end())
708 return offsets->end();
710 // This deals with cases (3) and (4).
711 if (offset == it->offset)
713 while (it != offsets->end()
714 && it->offset == offset
715 && it->line_num == -1)
717 if (it == offsets->end() || it->offset != offset)
718 return offsets->end();
723 // This handles the first part of case (7) -- we back up to the
724 // *first* entry that has the offset that's behind us.
725 gold_assert(it != offsets->begin());
726 std::vector<Offset_to_lineno_entry>::const_iterator range_end = it;
728 const off_t range_value = it->offset;
729 while (it != offsets->begin() && (it-1)->offset == range_value)
732 // This handles cases (5), (6), and (7): if any entry in the
733 // equal_range [it, range_end) has a line_num != -1, it's a valid
734 // match. If not, we're not in a function.
735 for (; it != range_end; ++it)
736 if (it->line_num != -1)
738 return offsets->end();
741 // Return a string for a file name and line number.
743 template<int size, bool big_endian>
745 Sized_dwarf_line_info<size, big_endian>::do_addr2line(unsigned int shndx,
748 if (this->data_valid_ == false)
751 const std::vector<Offset_to_lineno_entry>* offsets;
752 // If we do not have reloc information, then our input is a .so or
753 // some similar data structure where all the information is held in
754 // the offset. In that case, we ignore the input shndx.
755 if (this->input_is_relobj())
756 offsets = &this->line_number_map_[shndx];
758 offsets = &this->line_number_map_[-1U];
759 if (offsets->empty())
762 typename std::vector<Offset_to_lineno_entry>::const_iterator it
763 = offset_to_iterator(offsets, offset);
764 if (it == offsets->end())
767 // Convert the file_num + line_num into a string.
770 gold_assert(it->header_num < static_cast<int>(this->files_.size()));
771 gold_assert(it->file_num
772 < static_cast<int>(this->files_[it->header_num].size()));
773 const std::pair<int, std::string>& filename_pair
774 = this->files_[it->header_num][it->file_num];
775 const std::string& filename = filename_pair.second;
777 gold_assert(it->header_num < static_cast<int>(this->directories_.size()));
778 gold_assert(filename_pair.first
779 < static_cast<int>(this->directories_[it->header_num].size()));
780 const std::string& dirname
781 = this->directories_[it->header_num][filename_pair.first];
783 if (!dirname.empty())
792 char buffer[64]; // enough to hold a line number
793 snprintf(buffer, sizeof(buffer), "%d", it->line_num);
800 // Dwarf_line_info routines.
803 Dwarf_line_info::one_addr2line(Object* object,
804 unsigned int shndx, off_t offset)
806 switch (parameters->size_and_endianness())
808 #ifdef HAVE_TARGET_32_LITTLE
809 case Parameters::TARGET_32_LITTLE:
810 return Sized_dwarf_line_info<32, false>(object, shndx).addr2line(shndx,
813 #ifdef HAVE_TARGET_32_BIG
814 case Parameters::TARGET_32_BIG:
815 return Sized_dwarf_line_info<32, true>(object, shndx).addr2line(shndx,
818 #ifdef HAVE_TARGET_64_LITTLE
819 case Parameters::TARGET_64_LITTLE:
820 return Sized_dwarf_line_info<64, false>(object, shndx).addr2line(shndx,
823 #ifdef HAVE_TARGET_64_BIG
824 case Parameters::TARGET_64_BIG:
825 return Sized_dwarf_line_info<64, true>(object, shndx).addr2line(shndx,
833 #ifdef HAVE_TARGET_32_LITTLE
835 class Sized_dwarf_line_info<32, false>;
838 #ifdef HAVE_TARGET_32_BIG
840 class Sized_dwarf_line_info<32, true>;
843 #ifdef HAVE_TARGET_64_LITTLE
845 class Sized_dwarf_line_info<64, false>;
848 #ifdef HAVE_TARGET_64_BIG
850 class Sized_dwarf_line_info<64, true>;
853 } // End namespace gold.