1 // dwarf_reader.cc -- parse dwarf2/3 debug information
3 // Copyright 2007 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.
25 #include "elfcpp_swap.h"
29 #include "dwarf_reader.h"
33 // Read an unsigned LEB128 number. Each byte contains 7 bits of
34 // information, plus one bit saying whether the number continues or
38 read_unsigned_LEB_128(const unsigned char* buffer
, size_t* len
)
42 unsigned int shift
= 0;
49 result
|= (static_cast<uint64_t>(byte
& 0x7f)) << shift
;
59 // Read a signed LEB128 number. These are like regular LEB128
60 // numbers, except the last byte may have a sign bit set.
63 read_signed_LEB_128(const unsigned char* buffer
, size_t* len
)
74 result
|= (static_cast<uint64_t>(byte
& 0x7f) << shift
);
79 if ((shift
< 8 * static_cast<int>(sizeof(result
))) && (byte
& 0x40))
80 result
|= -((static_cast<int64_t>(1)) << shift
);
85 } // End anonymous namespace.
90 // This is the format of a DWARF2/3 line state machine that we process
91 // opcodes using. There is no need for anything outside the lineinfo
92 // processor to know how this works.
94 struct LineStateMachine
100 unsigned int shndx
; // the section address refers to
101 bool is_stmt
; // stmt means statement.
107 ResetLineStateMachine(struct LineStateMachine
* lsm
, bool default_is_stmt
)
114 lsm
->is_stmt
= default_is_stmt
;
115 lsm
->basic_block
= false;
116 lsm
->end_sequence
= false;
119 template<int size
, bool big_endian
>
120 Dwarf_line_info
<size
, big_endian
>::Dwarf_line_info(Object
* object
)
121 : data_valid_(false), buffer_(NULL
), symtab_buffer_(NULL
),
122 directories_(), files_(), current_header_index_(-1)
124 unsigned int debug_shndx
;
125 for (debug_shndx
= 0; debug_shndx
< object
->shnum(); ++debug_shndx
)
126 if (object
->section_name(debug_shndx
) == ".debug_line")
129 this->buffer_
= object
->section_contents(
130 debug_shndx
, &buffer_size
, false);
131 this->buffer_end_
= this->buffer_
+ buffer_size
;
134 if (this->buffer_
== NULL
)
137 // Find the relocation section for ".debug_line".
138 // We expect these for relobjs (.o's) but not dynobjs (.so's).
139 bool got_relocs
= false;
140 for (unsigned int reloc_shndx
= 0;
141 reloc_shndx
< object
->shnum();
144 unsigned int reloc_sh_type
= object
->section_type(reloc_shndx
);
145 if ((reloc_sh_type
== elfcpp::SHT_REL
146 || reloc_sh_type
== elfcpp::SHT_RELA
)
147 && object
->section_info(reloc_shndx
) == debug_shndx
)
149 got_relocs
= this->track_relocs_
.initialize(object
, reloc_shndx
,
155 // Finally, we need the symtab section to interpret the relocs.
158 unsigned int symtab_shndx
;
159 for (symtab_shndx
= 0; symtab_shndx
< object
->shnum(); ++symtab_shndx
)
160 if (object
->section_type(symtab_shndx
) == elfcpp::SHT_SYMTAB
)
162 this->symtab_buffer_
= object
->section_contents(
163 symtab_shndx
, &this->symtab_buffer_size_
, false);
166 if (this->symtab_buffer_
== NULL
)
170 // Now that we have successfully read all the data, parse the debug
172 this->data_valid_
= true;
173 this->read_line_mappings();
176 // Read the DWARF header.
178 template<int size
, bool big_endian
>
180 Dwarf_line_info
<size
, big_endian
>::read_header_prolog(
181 const unsigned char* lineptr
)
183 uint32_t initial_length
= elfcpp::Swap
<32, big_endian
>::readval(lineptr
);
186 // In DWARF2/3, if the initial length is all 1 bits, then the offset
187 // size is 8 and we need to read the next 8 bytes for the real length.
188 if (initial_length
== 0xffffffff)
190 header_
.offset_size
= 8;
191 initial_length
= elfcpp::Swap
<64, big_endian
>::readval(lineptr
);
195 header_
.offset_size
= 4;
197 header_
.total_length
= initial_length
;
199 gold_assert(lineptr
+ header_
.total_length
<= buffer_end_
);
201 header_
.version
= elfcpp::Swap
<16, big_endian
>::readval(lineptr
);
204 if (header_
.offset_size
== 4)
205 header_
.prologue_length
= elfcpp::Swap
<32, big_endian
>::readval(lineptr
);
207 header_
.prologue_length
= elfcpp::Swap
<64, big_endian
>::readval(lineptr
);
208 lineptr
+= header_
.offset_size
;
210 header_
.min_insn_length
= *lineptr
;
213 header_
.default_is_stmt
= *lineptr
;
216 header_
.line_base
= *reinterpret_cast<const signed char*>(lineptr
);
219 header_
.line_range
= *lineptr
;
222 header_
.opcode_base
= *lineptr
;
225 header_
.std_opcode_lengths
.reserve(header_
.opcode_base
+ 1);
226 header_
.std_opcode_lengths
[0] = 0;
227 for (int i
= 1; i
< header_
.opcode_base
; i
++)
229 header_
.std_opcode_lengths
[i
] = *lineptr
;
236 // The header for a debug_line section is mildly complicated, because
237 // the line info is very tightly encoded.
239 template<int size
, bool big_endian
>
241 Dwarf_line_info
<size
, big_endian
>::read_header_tables(
242 const unsigned char* lineptr
)
244 ++this->current_header_index_
;
246 // Create a new directories_ entry and a new files_ entry for our new
247 // header. We initialize each with a single empty element, because
248 // dwarf indexes directory and filenames starting at 1.
249 gold_assert(static_cast<int>(this->directories_
.size())
250 == this->current_header_index_
);
251 gold_assert(static_cast<int>(this->files_
.size())
252 == this->current_header_index_
);
253 this->directories_
.push_back(std::vector
<std::string
>(1));
254 this->files_
.push_back(std::vector
<std::pair
<int, std::string
> >(1));
256 // It is legal for the directory entry table to be empty.
262 const char* dirname
= reinterpret_cast<const char*>(lineptr
);
264 == static_cast<int>(this->directories_
.back().size()));
265 this->directories_
.back().push_back(dirname
);
266 lineptr
+= this->directories_
.back().back().size() + 1;
272 // It is also legal for the file entry table to be empty.
279 const char* filename
= reinterpret_cast<const char*>(lineptr
);
280 lineptr
+= strlen(filename
) + 1;
282 uint64_t dirindex
= read_unsigned_LEB_128(lineptr
, &len
);
285 if (dirindex
>= this->directories_
.back().size())
287 int dirindexi
= static_cast<int>(dirindex
);
289 read_unsigned_LEB_128(lineptr
, &len
); // mod_time
292 read_unsigned_LEB_128(lineptr
, &len
); // filelength
295 gold_assert(fileindex
296 == static_cast<int>(this->files_
.back().size()));
297 this->files_
.back().push_back(std::make_pair(dirindexi
, filename
));
306 // Process a single opcode in the .debug.line structure.
308 // Templating on size and big_endian would yield more efficient (and
309 // simpler) code, but would bloat the binary. Speed isn't important
312 template<int size
, bool big_endian
>
314 Dwarf_line_info
<size
, big_endian
>::process_one_opcode(
315 const unsigned char* start
, struct LineStateMachine
* lsm
, size_t* len
)
319 unsigned char opcode
= *start
;
323 // If the opcode is great than the opcode_base, it is a special
324 // opcode. Most line programs consist mainly of special opcodes.
325 if (opcode
>= header_
.opcode_base
)
327 opcode
-= header_
.opcode_base
;
328 const int advance_address
= ((opcode
/ header_
.line_range
)
329 * header_
.min_insn_length
);
330 lsm
->address
+= advance_address
;
332 const int advance_line
= ((opcode
% header_
.line_range
)
333 + header_
.line_base
);
334 lsm
->line_num
+= advance_line
;
335 lsm
->basic_block
= true;
340 // Otherwise, we have the regular opcodes
343 case elfcpp::DW_LNS_copy
:
344 lsm
->basic_block
= false;
348 case elfcpp::DW_LNS_advance_pc
:
350 const uint64_t advance_address
351 = read_unsigned_LEB_128(start
, &templen
);
353 lsm
->address
+= header_
.min_insn_length
* advance_address
;
357 case elfcpp::DW_LNS_advance_line
:
359 const uint64_t advance_line
= read_signed_LEB_128(start
, &templen
);
361 lsm
->line_num
+= advance_line
;
365 case elfcpp::DW_LNS_set_file
:
367 const uint64_t fileno
= read_unsigned_LEB_128(start
, &templen
);
369 lsm
->file_num
= fileno
;
373 case elfcpp::DW_LNS_set_column
:
375 const uint64_t colno
= read_unsigned_LEB_128(start
, &templen
);
377 lsm
->column_num
= colno
;
381 case elfcpp::DW_LNS_negate_stmt
:
382 lsm
->is_stmt
= !lsm
->is_stmt
;
385 case elfcpp::DW_LNS_set_basic_block
:
386 lsm
->basic_block
= true;
389 case elfcpp::DW_LNS_fixed_advance_pc
:
392 advance_address
= elfcpp::Swap
<16, big_endian
>::readval(start
);
394 lsm
->address
+= advance_address
;
398 case elfcpp::DW_LNS_const_add_pc
:
400 const int advance_address
= (header_
.min_insn_length
401 * ((255 - header_
.opcode_base
)
402 / header_
.line_range
));
403 lsm
->address
+= advance_address
;
407 case elfcpp::DW_LNS_extended_op
:
409 const uint64_t extended_op_len
410 = read_unsigned_LEB_128(start
, &templen
);
412 oplen
+= templen
+ extended_op_len
;
414 const unsigned char extended_op
= *start
;
419 case elfcpp::DW_LNE_end_sequence
:
420 lsm
->end_sequence
= true;
424 case elfcpp::DW_LNE_set_address
:
426 lsm
->address
= elfcpp::Swap
<size
, big_endian
>::readval(start
);
427 typename
Reloc_map::const_iterator it
428 = reloc_map_
.find(start
- this->buffer_
);
429 if (it
!= reloc_map_
.end())
432 lsm
->address
+= it
->second
.second
;
433 lsm
->shndx
= it
->second
.first
;
437 // If we're a normal .o file, with relocs, every
438 // set_address should have an associated relocation.
439 if (this->input_is_relobj())
440 this->data_valid_
= false;
444 case elfcpp::DW_LNE_define_file
:
446 const char* filename
= reinterpret_cast<const char*>(start
);
447 templen
= strlen(filename
) + 1;
450 uint64_t dirindex
= read_unsigned_LEB_128(start
, &templen
);
453 if (dirindex
>= this->directories_
.back().size())
455 int dirindexi
= static_cast<int>(dirindex
);
457 read_unsigned_LEB_128(start
, &templen
); // mod_time
460 read_unsigned_LEB_128(start
, &templen
); // filelength
463 this->files_
.back().push_back(std::make_pair(dirindexi
,
473 // Ignore unknown opcode silently
474 for (int i
= 0; i
< header_
.std_opcode_lengths
[opcode
]; i
++)
477 read_unsigned_LEB_128(start
, &templen
);
488 // Read the debug information at LINEPTR and store it in the line
491 template<int size
, bool big_endian
>
493 Dwarf_line_info
<size
, big_endian
>::read_lines(unsigned const char* lineptr
)
495 struct LineStateMachine lsm
;
497 // LENGTHSTART is the place the length field is based on. It is the
498 // point in the header after the initial length field.
499 const unsigned char* lengthstart
= buffer_
;
501 // In 64 bit dwarf, the initial length is 12 bytes, because of the
502 // 0xffffffff at the start.
503 if (header_
.offset_size
== 8)
508 while (lineptr
< lengthstart
+ header_
.total_length
)
510 ResetLineStateMachine(&lsm
, header_
.default_is_stmt
);
511 while (!lsm
.end_sequence
)
514 bool add_line
= this->process_one_opcode(lineptr
, &lsm
, &oplength
);
517 Offset_to_lineno_entry entry
518 = { lsm
.address
, this->current_header_index_
,
519 lsm
.file_num
, lsm
.line_num
};
520 line_number_map_
[lsm
.shndx
].push_back(entry
);
526 return lengthstart
+ header_
.total_length
;
529 // Looks in the symtab to see what section a symbol is in.
531 template<int size
, bool big_endian
>
533 Dwarf_line_info
<size
, big_endian
>::symbol_section(
535 typename
elfcpp::Elf_types
<size
>::Elf_Addr
* value
)
537 const int symsize
= elfcpp::Elf_sizes
<size
>::sym_size
;
538 gold_assert(sym
* symsize
< this->symtab_buffer_size_
);
539 elfcpp::Sym
<size
, big_endian
> elfsym(this->symtab_buffer_
+ sym
* symsize
);
540 *value
= elfsym
.get_st_value();
541 return elfsym
.get_st_shndx();
544 // Read the relocations into a Reloc_map.
546 template<int size
, bool big_endian
>
548 Dwarf_line_info
<size
, big_endian
>::read_relocs()
550 if (this->symtab_buffer_
== NULL
)
553 typename
elfcpp::Elf_types
<size
>::Elf_Addr value
;
555 while ((reloc_offset
= this->track_relocs_
.next_offset()) != -1)
557 const unsigned int sym
= this->track_relocs_
.next_symndx();
558 const unsigned int shndx
= this->symbol_section(sym
, &value
);
559 this->reloc_map_
[reloc_offset
] = std::make_pair(shndx
, value
);
560 this->track_relocs_
.advance(reloc_offset
+ 1);
564 // Read the line number info.
566 template<int size
, bool big_endian
>
568 Dwarf_line_info
<size
, big_endian
>::read_line_mappings()
570 gold_assert(this->data_valid_
== true);
573 while (this->buffer_
< this->buffer_end_
)
575 const unsigned char* lineptr
= this->buffer_
;
576 lineptr
= this->read_header_prolog(lineptr
);
577 lineptr
= this->read_header_tables(lineptr
);
578 lineptr
= this->read_lines(lineptr
);
579 this->buffer_
= lineptr
;
582 // Sort the lines numbers, so addr2line can use binary search.
583 for (typename
Lineno_map::iterator it
= line_number_map_
.begin();
584 it
!= line_number_map_
.end();
586 // Each vector needs to be sorted by offset.
587 std::sort(it
->second
.begin(), it
->second
.end());
590 // Some processing depends on whether the input is a .o file or not.
591 // For instance, .o files have relocs, and have .debug_lines
592 // information on a per section basis. .so files, on the other hand,
593 // lack relocs, and offsets are unique, so we can ignore the section
596 template<int size
, bool big_endian
>
598 Dwarf_line_info
<size
, big_endian
>::input_is_relobj()
600 // Only .o files have relocs and the symtab buffer that goes with them.
601 return this->symtab_buffer_
!= NULL
;
605 // Return a string for a file name and line number.
607 template<int size
, bool big_endian
>
609 Dwarf_line_info
<size
, big_endian
>::addr2line(unsigned int shndx
, off_t offset
)
611 if (this->data_valid_
== false)
614 const Offset_to_lineno_entry lookup_key
= { offset
, 0, 0, 0 };
615 const std::vector
<Offset_to_lineno_entry
>* offsets
;
616 // If we do not have reloc information, then our input is a .so or
617 // some similar data structure where all the information is held in
618 // the offset. In that case, we ignore the input shndx.
619 if (this->input_is_relobj())
620 offsets
= &this->line_number_map_
[shndx
];
622 offsets
= &this->line_number_map_
[-1U];
623 if (offsets
->empty())
626 typename
std::vector
<Offset_to_lineno_entry
>::const_iterator it
627 = std::lower_bound(offsets
->begin(), offsets
->end(), lookup_key
);
629 // If we found an exact match, great, otherwise find the last entry
630 // before the passed-in offset.
631 if (it
->offset
> offset
)
633 if (it
== offsets
->begin())
636 gold_assert(it
->offset
< offset
);
639 // Convert the file_num + line_num into a string.
642 gold_assert(it
->header_num
< static_cast<int>(this->files_
.size()));
643 gold_assert(it
->file_num
644 < static_cast<int>(this->files_
[it
->header_num
].size()));
645 const std::pair
<int, std::string
>& filename_pair
646 = this->files_
[it
->header_num
][it
->file_num
];
647 const std::string
& filename
= filename_pair
.second
;
649 gold_assert(it
->header_num
< static_cast<int>(this->directories_
.size()));
650 gold_assert(filename_pair
.first
651 < static_cast<int>(this->directories_
[it
->header_num
].size()));
652 const std::string
& dirname
653 = this->directories_
[it
->header_num
][filename_pair
.first
];
655 if (!dirname
.empty())
664 char buffer
[64]; // enough to hold a line number
665 snprintf(buffer
, sizeof(buffer
), "%d", it
->line_num
);
672 #ifdef HAVE_TARGET_32_LITTLE
674 class Dwarf_line_info
<32, false>;
677 #ifdef HAVE_TARGET_32_BIG
679 class Dwarf_line_info
<32, true>;
682 #ifdef HAVE_TARGET_64_LITTLE
684 class Dwarf_line_info
<64, false>;
687 #ifdef HAVE_TARGET_64_BIG
689 class Dwarf_line_info
<64, true>;
692 } // End namespace gold.