1 // dwarf_reader.cc -- parse dwarf2/3 debug information
3 // Copyright 2007 Free Software Foundation, Inc.
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.
25 #include "elfcpp_swap.h"
28 #include "parameters.h"
30 #include "dwarf_reader.h"
34 // Read an unsigned LEB128 number. Each byte contains 7 bits of
35 // information, plus one bit saying whether the number continues or
39 read_unsigned_LEB_128(const unsigned char* buffer, size_t* len)
43 unsigned int shift = 0;
50 result |= (static_cast<uint64_t>(byte & 0x7f)) << shift;
60 // Read a signed LEB128 number. These are like regular LEB128
61 // numbers, except the last byte may have a sign bit set.
64 read_signed_LEB_128(const unsigned char* buffer, size_t* len)
75 result |= (static_cast<uint64_t>(byte & 0x7f) << shift);
80 if ((shift < 8 * static_cast<int>(sizeof(result))) && (byte & 0x40))
81 result |= -((static_cast<int64_t>(1)) << shift);
86 } // End anonymous namespace.
91 // This is the format of a DWARF2/3 line state machine that we process
92 // opcodes using. There is no need for anything outside the lineinfo
93 // processor to know how this works.
95 struct LineStateMachine
101 unsigned int shndx; // the section address refers to
102 bool is_stmt; // stmt means statement.
108 ResetLineStateMachine(struct LineStateMachine* lsm, bool default_is_stmt)
115 lsm->is_stmt = default_is_stmt;
116 lsm->basic_block = false;
117 lsm->end_sequence = false;
120 template<int size, bool big_endian>
121 Sized_dwarf_line_info<size, big_endian>::Sized_dwarf_line_info(Object* object,
123 : data_valid_(false), buffer_(NULL), symtab_buffer_(NULL),
124 directories_(), files_(), current_header_index_(-1)
126 unsigned int debug_shndx;
127 for (debug_shndx = 0; debug_shndx < object->shnum(); ++debug_shndx)
128 if (object->section_name(debug_shndx) == ".debug_line")
131 this->buffer_ = object->section_contents(
132 debug_shndx, &buffer_size, false);
133 this->buffer_end_ = this->buffer_ + buffer_size;
136 if (this->buffer_ == NULL)
139 // Find the relocation section for ".debug_line".
140 // We expect these for relobjs (.o's) but not dynobjs (.so's).
141 bool got_relocs = false;
142 for (unsigned int reloc_shndx = 0;
143 reloc_shndx < object->shnum();
146 unsigned int reloc_sh_type = object->section_type(reloc_shndx);
147 if ((reloc_sh_type == elfcpp::SHT_REL
148 || reloc_sh_type == elfcpp::SHT_RELA)
149 && object->section_info(reloc_shndx) == debug_shndx)
151 got_relocs = this->track_relocs_.initialize(object, reloc_shndx,
157 // Finally, we need the symtab section to interpret the relocs.
160 unsigned int symtab_shndx;
161 for (symtab_shndx = 0; symtab_shndx < object->shnum(); ++symtab_shndx)
162 if (object->section_type(symtab_shndx) == elfcpp::SHT_SYMTAB)
164 this->symtab_buffer_ = object->section_contents(
165 symtab_shndx, &this->symtab_buffer_size_, false);
168 if (this->symtab_buffer_ == NULL)
172 // Now that we have successfully read all the data, parse the debug
174 this->data_valid_ = true;
175 this->read_line_mappings(read_shndx);
178 // Read the DWARF header.
180 template<int size, bool big_endian>
182 Sized_dwarf_line_info<size, big_endian>::read_header_prolog(
183 const unsigned char* lineptr)
185 uint32_t initial_length = elfcpp::Swap<32, big_endian>::readval(lineptr);
188 // In DWARF2/3, if the initial length is all 1 bits, then the offset
189 // size is 8 and we need to read the next 8 bytes for the real length.
190 if (initial_length == 0xffffffff)
192 header_.offset_size = 8;
193 initial_length = elfcpp::Swap<64, big_endian>::readval(lineptr);
197 header_.offset_size = 4;
199 header_.total_length = initial_length;
201 gold_assert(lineptr + header_.total_length <= buffer_end_);
203 header_.version = elfcpp::Swap<16, big_endian>::readval(lineptr);
206 if (header_.offset_size == 4)
207 header_.prologue_length = elfcpp::Swap<32, big_endian>::readval(lineptr);
209 header_.prologue_length = elfcpp::Swap<64, big_endian>::readval(lineptr);
210 lineptr += header_.offset_size;
212 header_.min_insn_length = *lineptr;
215 header_.default_is_stmt = *lineptr;
218 header_.line_base = *reinterpret_cast<const signed char*>(lineptr);
221 header_.line_range = *lineptr;
224 header_.opcode_base = *lineptr;
227 header_.std_opcode_lengths.reserve(header_.opcode_base + 1);
228 header_.std_opcode_lengths[0] = 0;
229 for (int i = 1; i < header_.opcode_base; i++)
231 header_.std_opcode_lengths[i] = *lineptr;
238 // The header for a debug_line section is mildly complicated, because
239 // the line info is very tightly encoded.
241 template<int size, bool big_endian>
243 Sized_dwarf_line_info<size, big_endian>::read_header_tables(
244 const unsigned char* lineptr)
246 ++this->current_header_index_;
248 // Create a new directories_ entry and a new files_ entry for our new
249 // header. We initialize each with a single empty element, because
250 // dwarf indexes directory and filenames starting at 1.
251 gold_assert(static_cast<int>(this->directories_.size())
252 == this->current_header_index_);
253 gold_assert(static_cast<int>(this->files_.size())
254 == this->current_header_index_);
255 this->directories_.push_back(std::vector<std::string>(1));
256 this->files_.push_back(std::vector<std::pair<int, std::string> >(1));
258 // It is legal for the directory entry table to be empty.
264 const char* dirname = reinterpret_cast<const char*>(lineptr);
266 == static_cast<int>(this->directories_.back().size()));
267 this->directories_.back().push_back(dirname);
268 lineptr += this->directories_.back().back().size() + 1;
274 // It is also legal for the file entry table to be empty.
281 const char* filename = reinterpret_cast<const char*>(lineptr);
282 lineptr += strlen(filename) + 1;
284 uint64_t dirindex = read_unsigned_LEB_128(lineptr, &len);
287 if (dirindex >= this->directories_.back().size())
289 int dirindexi = static_cast<int>(dirindex);
291 read_unsigned_LEB_128(lineptr, &len); // mod_time
294 read_unsigned_LEB_128(lineptr, &len); // filelength
297 gold_assert(fileindex
298 == static_cast<int>(this->files_.back().size()));
299 this->files_.back().push_back(std::make_pair(dirindexi, filename));
308 // Process a single opcode in the .debug.line structure.
310 // Templating on size and big_endian would yield more efficient (and
311 // simpler) code, but would bloat the binary. Speed isn't important
314 template<int size, bool big_endian>
316 Sized_dwarf_line_info<size, big_endian>::process_one_opcode(
317 const unsigned char* start, struct LineStateMachine* lsm, size_t* len)
321 unsigned char opcode = *start;
325 // If the opcode is great than the opcode_base, it is a special
326 // opcode. Most line programs consist mainly of special opcodes.
327 if (opcode >= header_.opcode_base)
329 opcode -= header_.opcode_base;
330 const int advance_address = ((opcode / header_.line_range)
331 * header_.min_insn_length);
332 lsm->address += advance_address;
334 const int advance_line = ((opcode % header_.line_range)
335 + header_.line_base);
336 lsm->line_num += advance_line;
337 lsm->basic_block = true;
342 // Otherwise, we have the regular opcodes
345 case elfcpp::DW_LNS_copy:
346 lsm->basic_block = false;
350 case elfcpp::DW_LNS_advance_pc:
352 const uint64_t advance_address
353 = read_unsigned_LEB_128(start, &templen);
355 lsm->address += header_.min_insn_length * advance_address;
359 case elfcpp::DW_LNS_advance_line:
361 const uint64_t advance_line = read_signed_LEB_128(start, &templen);
363 lsm->line_num += advance_line;
367 case elfcpp::DW_LNS_set_file:
369 const uint64_t fileno = read_unsigned_LEB_128(start, &templen);
371 lsm->file_num = fileno;
375 case elfcpp::DW_LNS_set_column:
377 const uint64_t colno = read_unsigned_LEB_128(start, &templen);
379 lsm->column_num = colno;
383 case elfcpp::DW_LNS_negate_stmt:
384 lsm->is_stmt = !lsm->is_stmt;
387 case elfcpp::DW_LNS_set_basic_block:
388 lsm->basic_block = true;
391 case elfcpp::DW_LNS_fixed_advance_pc:
394 advance_address = elfcpp::Swap<16, big_endian>::readval(start);
396 lsm->address += advance_address;
400 case elfcpp::DW_LNS_const_add_pc:
402 const int advance_address = (header_.min_insn_length
403 * ((255 - header_.opcode_base)
404 / header_.line_range));
405 lsm->address += advance_address;
409 case elfcpp::DW_LNS_extended_op:
411 const uint64_t extended_op_len
412 = read_unsigned_LEB_128(start, &templen);
414 oplen += templen + extended_op_len;
416 const unsigned char extended_op = *start;
421 case elfcpp::DW_LNE_end_sequence:
422 // This means that the current byte is the one immediately
423 // after a set of instructions. Record the current line
424 // for up to one less than the current address.
426 lsm->end_sequence = true;
430 case elfcpp::DW_LNE_set_address:
432 lsm->address = elfcpp::Swap<size, big_endian>::readval(start);
433 typename Reloc_map::const_iterator it
434 = reloc_map_.find(start - this->buffer_);
435 if (it != reloc_map_.end())
438 lsm->address += it->second.second;
439 lsm->shndx = it->second.first;
443 // If we're a normal .o file, with relocs, every
444 // set_address should have an associated relocation.
445 if (this->input_is_relobj())
446 this->data_valid_ = false;
450 case elfcpp::DW_LNE_define_file:
452 const char* filename = reinterpret_cast<const char*>(start);
453 templen = strlen(filename) + 1;
456 uint64_t dirindex = read_unsigned_LEB_128(start, &templen);
459 if (dirindex >= this->directories_.back().size())
461 int dirindexi = static_cast<int>(dirindex);
463 read_unsigned_LEB_128(start, &templen); // mod_time
466 read_unsigned_LEB_128(start, &templen); // filelength
469 this->files_.back().push_back(std::make_pair(dirindexi,
479 // Ignore unknown opcode silently
480 for (int i = 0; i < header_.std_opcode_lengths[opcode]; i++)
483 read_unsigned_LEB_128(start, &templen);
494 // Read the debug information at LINEPTR and store it in the line
497 template<int size, bool big_endian>
499 Sized_dwarf_line_info<size, big_endian>::read_lines(unsigned const char* lineptr,
502 struct LineStateMachine lsm;
504 // LENGTHSTART is the place the length field is based on. It is the
505 // point in the header after the initial length field.
506 const unsigned char* lengthstart = buffer_;
508 // In 64 bit dwarf, the initial length is 12 bytes, because of the
509 // 0xffffffff at the start.
510 if (header_.offset_size == 8)
515 while (lineptr < lengthstart + header_.total_length)
517 ResetLineStateMachine(&lsm, header_.default_is_stmt);
518 while (!lsm.end_sequence)
521 bool add_line = this->process_one_opcode(lineptr, &lsm, &oplength);
523 && (shndx == -1U || lsm.shndx == -1U || shndx == lsm.shndx))
525 Offset_to_lineno_entry entry
526 = { lsm.address, this->current_header_index_,
527 lsm.file_num, lsm.line_num };
528 line_number_map_[lsm.shndx].push_back(entry);
534 return lengthstart + header_.total_length;
537 // Looks in the symtab to see what section a symbol is in.
539 template<int size, bool big_endian>
541 Sized_dwarf_line_info<size, big_endian>::symbol_section(
543 typename elfcpp::Elf_types<size>::Elf_Addr* value)
545 const int symsize = elfcpp::Elf_sizes<size>::sym_size;
546 gold_assert(sym * symsize < this->symtab_buffer_size_);
547 elfcpp::Sym<size, big_endian> elfsym(this->symtab_buffer_ + sym * symsize);
548 *value = elfsym.get_st_value();
549 return elfsym.get_st_shndx();
552 // Read the relocations into a Reloc_map.
554 template<int size, bool big_endian>
556 Sized_dwarf_line_info<size, big_endian>::read_relocs()
558 if (this->symtab_buffer_ == NULL)
561 typename elfcpp::Elf_types<size>::Elf_Addr value;
563 while ((reloc_offset = this->track_relocs_.next_offset()) != -1)
565 const unsigned int sym = this->track_relocs_.next_symndx();
566 const unsigned int shndx = this->symbol_section(sym, &value);
567 this->reloc_map_[reloc_offset] = std::make_pair(shndx, value);
568 this->track_relocs_.advance(reloc_offset + 1);
572 // Read the line number info.
574 template<int size, bool big_endian>
576 Sized_dwarf_line_info<size, big_endian>::read_line_mappings(off_t shndx)
578 gold_assert(this->data_valid_ == true);
581 while (this->buffer_ < this->buffer_end_)
583 const unsigned char* lineptr = this->buffer_;
584 lineptr = this->read_header_prolog(lineptr);
585 lineptr = this->read_header_tables(lineptr);
586 lineptr = this->read_lines(lineptr, shndx);
587 this->buffer_ = lineptr;
590 // Sort the lines numbers, so addr2line can use binary search.
591 for (typename Lineno_map::iterator it = line_number_map_.begin();
592 it != line_number_map_.end();
594 // Each vector needs to be sorted by offset.
595 std::sort(it->second.begin(), it->second.end());
598 // Some processing depends on whether the input is a .o file or not.
599 // For instance, .o files have relocs, and have .debug_lines
600 // information on a per section basis. .so files, on the other hand,
601 // lack relocs, and offsets are unique, so we can ignore the section
604 template<int size, bool big_endian>
606 Sized_dwarf_line_info<size, big_endian>::input_is_relobj()
608 // Only .o files have relocs and the symtab buffer that goes with them.
609 return this->symtab_buffer_ != NULL;
612 // Given an Offset_to_lineno_entry vector, and an offset, figure out
613 // if the offset points into a function according to the vector (see
614 // comments below for the algorithm). If it does, return an iterator
615 // into the vector that points to the line-number that contains that
616 // offset. If not, it returns vector::end().
618 static std::vector<Offset_to_lineno_entry>::const_iterator
619 offset_to_iterator(const std::vector<Offset_to_lineno_entry>* offsets,
622 const Offset_to_lineno_entry lookup_key = { offset, 0, 0, 0 };
624 // lower_bound() returns the smallest offset which is >= lookup_key.
625 // If no offset in offsets is >= lookup_key, returns end().
626 std::vector<Offset_to_lineno_entry>::const_iterator it
627 = std::lower_bound(offsets->begin(), offsets->end(), lookup_key);
629 // This code is easiest to understand with a concrete example.
630 // Here's a possible offsets array:
631 // {{offset = 3211, header_num = 0, file_num = 1, line_num = 16}, // 0
632 // {offset = 3224, header_num = 0, file_num = 1, line_num = 20}, // 1
633 // {offset = 3226, header_num = 0, file_num = 1, line_num = 22}, // 2
634 // {offset = 3231, header_num = 0, file_num = 1, line_num = 25}, // 3
635 // {offset = 3232, header_num = 0, file_num = 1, line_num = -1}, // 4
636 // {offset = 3232, header_num = 0, file_num = 1, line_num = 65}, // 5
637 // {offset = 3235, header_num = 0, file_num = 1, line_num = 66}, // 6
638 // {offset = 3236, header_num = 0, file_num = 1, line_num = -1}, // 7
639 // {offset = 5764, header_num = 0, file_num = 1, line_num = 47}, // 8
640 // {offset = 5765, header_num = 0, file_num = 1, line_num = 48}, // 9
641 // {offset = 5767, header_num = 0, file_num = 1, line_num = 49}, // 10
642 // {offset = 5768, header_num = 0, file_num = 1, line_num = 50}, // 11
643 // {offset = 5773, header_num = 0, file_num = 1, line_num = -1}, // 12
644 // {offset = 5787, header_num = 1, file_num = 1, line_num = 19}, // 13
645 // {offset = 5790, header_num = 1, file_num = 1, line_num = 20}, // 14
646 // {offset = 5793, header_num = 1, file_num = 1, line_num = 67}, // 15
647 // {offset = 5793, header_num = 1, file_num = 1, line_num = -1}, // 16
648 // {offset = 5795, header_num = 1, file_num = 1, line_num = 68}, // 17
649 // {offset = 5798, header_num = 1, file_num = 1, line_num = -1}, // 18
650 // The entries with line_num == -1 mark the end of a function: the
651 // associated offset is one past the last instruction in the
652 // function. This can correspond to the beginning of the next
653 // function (as is true for offset 3232); alternately, there can be
654 // a gap between the end of one function and the start of the next
655 // (as is true for some others, most obviously from 3236->5764).
657 // Case 1: lookup_key has offset == 10. lower_bound returns
658 // offsets[0]. Since it's not an exact match and we're
659 // at the beginning of offsets, we return end() (invalid).
660 // Case 2: lookup_key has offset 10000. lower_bound returns
661 // offset[19] (end()). We return end() (invalid).
662 // Case 3: lookup_key has offset == 3211. lower_bound matches
663 // offsets[0] exactly, and that's the entry we return.
664 // Case 4: lookup_key has offset == 3232. lower_bound returns
665 // offsets[4]. That's an exact match, but indicates
666 // end-of-function. We check if offsets[5] is also an
667 // exact match but not end-of-function. It is, so we
668 // return offsets[5].
669 // Case 5: lookup_key has offset == 3214. lower_bound returns
670 // offsets[1]. Since it's not an exact match, we back
671 // up to the offset that's < lookup_key, offsets[0].
672 // We note offsets[0] is a valid entry (not end-of-function),
673 // so that's the entry we return.
674 // Case 6: lookup_key has offset == 4000. lower_bound returns
675 // offsets[8]. Since it's not an exact match, we back
676 // up to offsets[7]. Since offsets[7] indicates
677 // end-of-function, we know lookup_key is between
678 // functions, so we return end() (not a valid offset).
679 // Case 7: lookup_key has offset == 5794. lower_bound returns
680 // offsets[17]. Since it's not an exact match, we back
681 // up to offsets[15]. Note we back up to the *first*
682 // entry with offset 5793, not just offsets[17-1].
683 // We note offsets[15] is a valid entry, so we return it.
684 // If offsets[15] had had line_num == -1, we would have
685 // checked offsets[16]. The reason for this is that
686 // 15 and 16 can be in an arbitrary order, since we sort
687 // only by offset. (Note it doesn't help to use line_number
688 // as a secondary sort key, since sometimes we want the -1
689 // to be first and sometimes we want it to be last.)
691 // This deals with cases (1) and (2).
692 if ((it == offsets->begin() && offset < it->offset)
693 || it == offsets->end())
694 return offsets->end();
696 // This deals with cases (3) and (4).
697 if (offset == it->offset)
699 while (it != offsets->end()
700 && it->offset == offset
701 && it->line_num == -1)
703 if (it == offsets->end() || it->offset != offset)
704 return offsets->end();
709 // This handles the first part of case (7) -- we back up to the
710 // *first* entry that has the offset that's behind us.
711 gold_assert(it != offsets->begin());
712 std::vector<Offset_to_lineno_entry>::const_iterator range_end = it;
714 const off_t range_value = it->offset;
715 while (it != offsets->begin() && (it-1)->offset == range_value)
718 // This handles cases (5), (6), and (7): if any entry in the
719 // equal_range [it, range_end) has a line_num != -1, it's a valid
720 // match. If not, we're not in a function.
721 for (; it != range_end; ++it)
722 if (it->line_num != -1)
724 return offsets->end();
727 // Return a string for a file name and line number.
729 template<int size, bool big_endian>
731 Sized_dwarf_line_info<size, big_endian>::do_addr2line(unsigned int shndx,
734 if (this->data_valid_ == false)
737 const std::vector<Offset_to_lineno_entry>* offsets;
738 // If we do not have reloc information, then our input is a .so or
739 // some similar data structure where all the information is held in
740 // the offset. In that case, we ignore the input shndx.
741 if (this->input_is_relobj())
742 offsets = &this->line_number_map_[shndx];
744 offsets = &this->line_number_map_[-1U];
745 if (offsets->empty())
748 typename std::vector<Offset_to_lineno_entry>::const_iterator it
749 = offset_to_iterator(offsets, offset);
750 if (it == offsets->end())
753 // Convert the file_num + line_num into a string.
756 gold_assert(it->header_num < static_cast<int>(this->files_.size()));
757 gold_assert(it->file_num
758 < static_cast<int>(this->files_[it->header_num].size()));
759 const std::pair<int, std::string>& filename_pair
760 = this->files_[it->header_num][it->file_num];
761 const std::string& filename = filename_pair.second;
763 gold_assert(it->header_num < static_cast<int>(this->directories_.size()));
764 gold_assert(filename_pair.first
765 < static_cast<int>(this->directories_[it->header_num].size()));
766 const std::string& dirname
767 = this->directories_[it->header_num][filename_pair.first];
769 if (!dirname.empty())
778 char buffer[64]; // enough to hold a line number
779 snprintf(buffer, sizeof(buffer), "%d", it->line_num);
786 // Dwarf_line_info routines.
789 Dwarf_line_info::one_addr2line(Object* object,
790 unsigned int shndx, off_t offset)
792 if (parameters->get_size() == 32 && !parameters->is_big_endian())
793 #ifdef HAVE_TARGET_32_LITTLE
794 return Sized_dwarf_line_info<32, false>(object, shndx).addr2line(shndx,
799 else if (parameters->get_size() == 32 && parameters->is_big_endian())
800 #ifdef HAVE_TARGET_32_BIG
801 return Sized_dwarf_line_info<32, true>(object, shndx).addr2line(shndx,
806 else if (parameters->get_size() == 64 && !parameters->is_big_endian())
807 #ifdef HAVE_TARGET_64_LITTLE
808 return Sized_dwarf_line_info<64, false>(object, shndx).addr2line(shndx,
813 else if (parameters->get_size() == 64 && parameters->is_big_endian())
814 #ifdef HAVE_TARGET_64_BIT
815 return Sized_dwarf_line_info<64, true>(object, shndx).addr2line(shndx,
824 #ifdef HAVE_TARGET_32_LITTLE
826 class Sized_dwarf_line_info<32, false>;
829 #ifdef HAVE_TARGET_32_BIG
831 class Sized_dwarf_line_info<32, true>;
834 #ifdef HAVE_TARGET_64_LITTLE
836 class Sized_dwarf_line_info<64, false>;
839 #ifdef HAVE_TARGET_64_BIG
841 class Sized_dwarf_line_info<64, true>;
844 } // End namespace gold.