// dwarf_reader.cc -- parse dwarf2/3 debug information
-// Copyright 2007 Free Software Foundation, Inc.
+// Copyright 2007, 2008, 2009 Free Software Foundation, Inc.
// Written by Ian Lance Taylor <iant@google.com>.
// This file is part of gold.
#include "gold.h"
+#include <algorithm>
+#include <vector>
+
#include "elfcpp_swap.h"
#include "dwarf.h"
#include "object.h"
+#include "parameters.h"
#include "reloc.h"
#include "dwarf_reader.h"
-
-namespace {
-
-// Read an unsigned LEB128 number. Each byte contains 7 bits of
-// information, plus one bit saying whether the number continues or
-// not.
-
-uint64_t
-read_unsigned_LEB_128(const unsigned char* buffer, size_t* len)
-{
- uint64_t result = 0;
- size_t num_read = 0;
- unsigned int shift = 0;
- unsigned char byte;
-
- do
- {
- byte = *buffer++;
- num_read++;
- result |= (static_cast<uint64_t>(byte & 0x7f)) << shift;
- shift += 7;
- }
- while (byte & 0x80);
-
- *len = num_read;
-
- return result;
-}
-
-// Read a signed LEB128 number. These are like regular LEB128
-// numbers, except the last byte may have a sign bit set.
-
-int64_t
-read_signed_LEB_128(const unsigned char* buffer, size_t* len)
-{
- int64_t result = 0;
- int shift = 0;
- size_t num_read = 0;
- unsigned char byte;
-
- do
- {
- byte = *buffer++;
- num_read++;
- result |= (static_cast<uint64_t>(byte & 0x7f) << shift);
- shift += 7;
- }
- while (byte & 0x80);
-
- if ((shift < 8 * static_cast<int>(sizeof(result))) && (byte & 0x40))
- result |= -((static_cast<int64_t>(1)) << shift);
- *len = num_read;
- return result;
-}
-
-} // End anonymous namespace.
-
+#include "int_encoding.h"
+#include "compressed_output.h"
namespace gold {
-// This is the format of a DWARF2/3 line state machine that we process
-// opcodes using. There is no need for anything outside the lineinfo
-// processor to know how this works.
-
struct LineStateMachine
{
int file_num;
}
template<int size, bool big_endian>
-Dwarf_line_info<size, big_endian>::Dwarf_line_info(Object* object)
+Sized_dwarf_line_info<size, big_endian>::Sized_dwarf_line_info(Object* object,
+ unsigned int read_shndx)
: data_valid_(false), buffer_(NULL), symtab_buffer_(NULL),
directories_(), files_(), current_header_index_(-1)
{
unsigned int debug_shndx;
for (debug_shndx = 0; debug_shndx < object->shnum(); ++debug_shndx)
+ // FIXME: do this more efficiently: section_name() isn't super-fast
if (object->section_name(debug_shndx) == ".debug_line")
{
- off_t buffer_size;
- this->buffer_ = object->section_contents(
- debug_shndx, &buffer_size, false);
+ section_size_type buffer_size;
+ this->buffer_ = object->section_contents(debug_shndx, &buffer_size,
+ false);
this->buffer_end_ = this->buffer_ + buffer_size;
break;
}
if (this->buffer_ == NULL)
return;
+ section_size_type uncompressed_size = 0;
+ unsigned char* uncompressed_data = NULL;
+ if (object->section_is_compressed(debug_shndx, &uncompressed_size))
+ {
+ uncompressed_data = new unsigned char[uncompressed_size];
+ if (!decompress_input_section(this->buffer_,
+ this->buffer_end_ - this->buffer_,
+ uncompressed_data,
+ uncompressed_size))
+ object->error(_("could not decompress section %s"),
+ object->section_name(debug_shndx).c_str());
+ this->buffer_ = uncompressed_data;
+ this->buffer_end_ = this->buffer_ + uncompressed_size;
+ }
+
// Find the relocation section for ".debug_line".
// We expect these for relobjs (.o's) but not dynobjs (.so's).
bool got_relocs = false;
// Now that we have successfully read all the data, parse the debug
// info.
this->data_valid_ = true;
- this->read_line_mappings();
+ this->read_line_mappings(object, read_shndx);
}
// Read the DWARF header.
template<int size, bool big_endian>
const unsigned char*
-Dwarf_line_info<size, big_endian>::read_header_prolog(
+Sized_dwarf_line_info<size, big_endian>::read_header_prolog(
const unsigned char* lineptr)
{
- uint32_t initial_length = elfcpp::Swap<32, big_endian>::readval(lineptr);
+ uint32_t initial_length = elfcpp::Swap_unaligned<32, big_endian>::readval(lineptr);
lineptr += 4;
// In DWARF2/3, if the initial length is all 1 bits, then the offset
if (initial_length == 0xffffffff)
{
header_.offset_size = 8;
- initial_length = elfcpp::Swap<64, big_endian>::readval(lineptr);
+ initial_length = elfcpp::Swap_unaligned<64, big_endian>::readval(lineptr);
lineptr += 8;
}
else
gold_assert(lineptr + header_.total_length <= buffer_end_);
- header_.version = elfcpp::Swap<16, big_endian>::readval(lineptr);
+ header_.version = elfcpp::Swap_unaligned<16, big_endian>::readval(lineptr);
lineptr += 2;
if (header_.offset_size == 4)
- header_.prologue_length = elfcpp::Swap<32, big_endian>::readval(lineptr);
+ header_.prologue_length = elfcpp::Swap_unaligned<32, big_endian>::readval(lineptr);
else
- header_.prologue_length = elfcpp::Swap<64, big_endian>::readval(lineptr);
+ header_.prologue_length = elfcpp::Swap_unaligned<64, big_endian>::readval(lineptr);
lineptr += header_.offset_size;
header_.min_insn_length = *lineptr;
template<int size, bool big_endian>
const unsigned char*
-Dwarf_line_info<size, big_endian>::read_header_tables(
+Sized_dwarf_line_info<size, big_endian>::read_header_tables(
const unsigned char* lineptr)
{
++this->current_header_index_;
template<int size, bool big_endian>
bool
-Dwarf_line_info<size, big_endian>::process_one_opcode(
+Sized_dwarf_line_info<size, big_endian>::process_one_opcode(
const unsigned char* start, struct LineStateMachine* lsm, size_t* len)
{
size_t oplen = 0;
case elfcpp::DW_LNS_fixed_advance_pc:
{
int advance_address;
- advance_address = elfcpp::Swap<16, big_endian>::readval(start);
+ advance_address = elfcpp::Swap_unaligned<16, big_endian>::readval(start);
oplen += 2;
lsm->address += advance_address;
}
switch (extended_op)
{
case elfcpp::DW_LNE_end_sequence:
+ // This means that the current byte is the one immediately
+ // after a set of instructions. Record the current line
+ // for up to one less than the current address.
+ lsm->line_num = -1;
lsm->end_sequence = true;
*len = oplen;
return true;
case elfcpp::DW_LNE_set_address:
{
- lsm->address = elfcpp::Swap<size, big_endian>::readval(start);
+ lsm->address = elfcpp::Swap_unaligned<size, big_endian>::readval(start);
typename Reloc_map::const_iterator it
= reloc_map_.find(start - this->buffer_);
if (it != reloc_map_.end())
template<int size, bool big_endian>
unsigned const char*
-Dwarf_line_info<size, big_endian>::read_lines(unsigned const char* lineptr)
+Sized_dwarf_line_info<size, big_endian>::read_lines(unsigned const char* lineptr,
+ unsigned int shndx)
{
struct LineStateMachine lsm;
{
size_t oplength;
bool add_line = this->process_one_opcode(lineptr, &lsm, &oplength);
- if (add_line)
+ if (add_line
+ && (shndx == -1U || lsm.shndx == -1U || shndx == lsm.shndx))
{
Offset_to_lineno_entry entry
= { lsm.address, this->current_header_index_,
template<int size, bool big_endian>
unsigned int
-Dwarf_line_info<size, big_endian>::symbol_section(
+Sized_dwarf_line_info<size, big_endian>::symbol_section(
+ Object* object,
unsigned int sym,
- typename elfcpp::Elf_types<size>::Elf_Addr* value)
+ typename elfcpp::Elf_types<size>::Elf_Addr* value,
+ bool* is_ordinary)
{
const int symsize = elfcpp::Elf_sizes<size>::sym_size;
gold_assert(sym * symsize < this->symtab_buffer_size_);
elfcpp::Sym<size, big_endian> elfsym(this->symtab_buffer_ + sym * symsize);
*value = elfsym.get_st_value();
- return elfsym.get_st_shndx();
+ return object->adjust_sym_shndx(sym, elfsym.get_st_shndx(), is_ordinary);
}
// Read the relocations into a Reloc_map.
template<int size, bool big_endian>
void
-Dwarf_line_info<size, big_endian>::read_relocs()
+Sized_dwarf_line_info<size, big_endian>::read_relocs(Object* object)
{
if (this->symtab_buffer_ == NULL)
return;
while ((reloc_offset = this->track_relocs_.next_offset()) != -1)
{
const unsigned int sym = this->track_relocs_.next_symndx();
- const unsigned int shndx = this->symbol_section(sym, &value);
- this->reloc_map_[reloc_offset] = std::make_pair(shndx, value);
+
+ bool is_ordinary;
+ const unsigned int shndx = this->symbol_section(object, sym, &value,
+ &is_ordinary);
+
+ // There is no reason to record non-ordinary section indexes, or
+ // SHN_UNDEF, because they will never match the real section.
+ if (is_ordinary && shndx != elfcpp::SHN_UNDEF)
+ this->reloc_map_[reloc_offset] = std::make_pair(shndx, value);
+
this->track_relocs_.advance(reloc_offset + 1);
}
}
template<int size, bool big_endian>
void
-Dwarf_line_info<size, big_endian>::read_line_mappings()
+Sized_dwarf_line_info<size, big_endian>::read_line_mappings(Object* object,
+ unsigned int shndx)
{
gold_assert(this->data_valid_ == true);
- read_relocs();
+ this->read_relocs(object);
while (this->buffer_ < this->buffer_end_)
{
const unsigned char* lineptr = this->buffer_;
lineptr = this->read_header_prolog(lineptr);
lineptr = this->read_header_tables(lineptr);
- lineptr = this->read_lines(lineptr);
+ lineptr = this->read_lines(lineptr, shndx);
this->buffer_ = lineptr;
}
template<int size, bool big_endian>
bool
-Dwarf_line_info<size, big_endian>::input_is_relobj()
+Sized_dwarf_line_info<size, big_endian>::input_is_relobj()
{
// Only .o files have relocs and the symtab buffer that goes with them.
return this->symtab_buffer_ != NULL;
}
+// Given an Offset_to_lineno_entry vector, and an offset, figure out
+// if the offset points into a function according to the vector (see
+// comments below for the algorithm). If it does, return an iterator
+// into the vector that points to the line-number that contains that
+// offset. If not, it returns vector::end().
+
+static std::vector<Offset_to_lineno_entry>::const_iterator
+offset_to_iterator(const std::vector<Offset_to_lineno_entry>* offsets,
+ off_t offset)
+{
+ const Offset_to_lineno_entry lookup_key = { offset, 0, 0, 0 };
+
+ // lower_bound() returns the smallest offset which is >= lookup_key.
+ // If no offset in offsets is >= lookup_key, returns end().
+ std::vector<Offset_to_lineno_entry>::const_iterator it
+ = std::lower_bound(offsets->begin(), offsets->end(), lookup_key);
+
+ // This code is easiest to understand with a concrete example.
+ // Here's a possible offsets array:
+ // {{offset = 3211, header_num = 0, file_num = 1, line_num = 16}, // 0
+ // {offset = 3224, header_num = 0, file_num = 1, line_num = 20}, // 1
+ // {offset = 3226, header_num = 0, file_num = 1, line_num = 22}, // 2
+ // {offset = 3231, header_num = 0, file_num = 1, line_num = 25}, // 3
+ // {offset = 3232, header_num = 0, file_num = 1, line_num = -1}, // 4
+ // {offset = 3232, header_num = 0, file_num = 1, line_num = 65}, // 5
+ // {offset = 3235, header_num = 0, file_num = 1, line_num = 66}, // 6
+ // {offset = 3236, header_num = 0, file_num = 1, line_num = -1}, // 7
+ // {offset = 5764, header_num = 0, file_num = 1, line_num = 47}, // 8
+ // {offset = 5765, header_num = 0, file_num = 1, line_num = 48}, // 9
+ // {offset = 5767, header_num = 0, file_num = 1, line_num = 49}, // 10
+ // {offset = 5768, header_num = 0, file_num = 1, line_num = 50}, // 11
+ // {offset = 5773, header_num = 0, file_num = 1, line_num = -1}, // 12
+ // {offset = 5787, header_num = 1, file_num = 1, line_num = 19}, // 13
+ // {offset = 5790, header_num = 1, file_num = 1, line_num = 20}, // 14
+ // {offset = 5793, header_num = 1, file_num = 1, line_num = 67}, // 15
+ // {offset = 5793, header_num = 1, file_num = 1, line_num = -1}, // 16
+ // {offset = 5795, header_num = 1, file_num = 1, line_num = 68}, // 17
+ // {offset = 5798, header_num = 1, file_num = 1, line_num = -1}, // 18
+ // The entries with line_num == -1 mark the end of a function: the
+ // associated offset is one past the last instruction in the
+ // function. This can correspond to the beginning of the next
+ // function (as is true for offset 3232); alternately, there can be
+ // a gap between the end of one function and the start of the next
+ // (as is true for some others, most obviously from 3236->5764).
+ //
+ // Case 1: lookup_key has offset == 10. lower_bound returns
+ // offsets[0]. Since it's not an exact match and we're
+ // at the beginning of offsets, we return end() (invalid).
+ // Case 2: lookup_key has offset 10000. lower_bound returns
+ // offset[19] (end()). We return end() (invalid).
+ // Case 3: lookup_key has offset == 3211. lower_bound matches
+ // offsets[0] exactly, and that's the entry we return.
+ // Case 4: lookup_key has offset == 3232. lower_bound returns
+ // offsets[4]. That's an exact match, but indicates
+ // end-of-function. We check if offsets[5] is also an
+ // exact match but not end-of-function. It is, so we
+ // return offsets[5].
+ // Case 5: lookup_key has offset == 3214. lower_bound returns
+ // offsets[1]. Since it's not an exact match, we back
+ // up to the offset that's < lookup_key, offsets[0].
+ // We note offsets[0] is a valid entry (not end-of-function),
+ // so that's the entry we return.
+ // Case 6: lookup_key has offset == 4000. lower_bound returns
+ // offsets[8]. Since it's not an exact match, we back
+ // up to offsets[7]. Since offsets[7] indicates
+ // end-of-function, we know lookup_key is between
+ // functions, so we return end() (not a valid offset).
+ // Case 7: lookup_key has offset == 5794. lower_bound returns
+ // offsets[17]. Since it's not an exact match, we back
+ // up to offsets[15]. Note we back up to the *first*
+ // entry with offset 5793, not just offsets[17-1].
+ // We note offsets[15] is a valid entry, so we return it.
+ // If offsets[15] had had line_num == -1, we would have
+ // checked offsets[16]. The reason for this is that
+ // 15 and 16 can be in an arbitrary order, since we sort
+ // only by offset. (Note it doesn't help to use line_number
+ // as a secondary sort key, since sometimes we want the -1
+ // to be first and sometimes we want it to be last.)
+
+ // This deals with cases (1) and (2).
+ if ((it == offsets->begin() && offset < it->offset)
+ || it == offsets->end())
+ return offsets->end();
+
+ // This deals with cases (3) and (4).
+ if (offset == it->offset)
+ {
+ while (it != offsets->end()
+ && it->offset == offset
+ && it->line_num == -1)
+ ++it;
+ if (it == offsets->end() || it->offset != offset)
+ return offsets->end();
+ else
+ return it;
+ }
+
+ // This handles the first part of case (7) -- we back up to the
+ // *first* entry that has the offset that's behind us.
+ gold_assert(it != offsets->begin());
+ std::vector<Offset_to_lineno_entry>::const_iterator range_end = it;
+ --it;
+ const off_t range_value = it->offset;
+ while (it != offsets->begin() && (it-1)->offset == range_value)
+ --it;
+
+ // This handles cases (5), (6), and (7): if any entry in the
+ // equal_range [it, range_end) has a line_num != -1, it's a valid
+ // match. If not, we're not in a function.
+ for (; it != range_end; ++it)
+ if (it->line_num != -1)
+ return it;
+ return offsets->end();
+}
// Return a string for a file name and line number.
template<int size, bool big_endian>
std::string
-Dwarf_line_info<size, big_endian>::addr2line(unsigned int shndx, off_t offset)
+Sized_dwarf_line_info<size, big_endian>::do_addr2line(unsigned int shndx,
+ off_t offset)
{
if (this->data_valid_ == false)
return "";
- const Offset_to_lineno_entry lookup_key = { offset, 0, 0, 0 };
const std::vector<Offset_to_lineno_entry>* offsets;
// If we do not have reloc information, then our input is a .so or
// some similar data structure where all the information is held in
return "";
typename std::vector<Offset_to_lineno_entry>::const_iterator it
- = std::lower_bound(offsets->begin(), offsets->end(), lookup_key);
-
- // If we found an exact match, great, otherwise find the last entry
- // before the passed-in offset.
- if (it->offset > offset)
- {
- if (it == offsets->begin())
- return "";
- --it;
- gold_assert(it->offset < offset);
- }
+ = offset_to_iterator(offsets, offset);
+ if (it == offsets->end())
+ return "";
// Convert the file_num + line_num into a string.
std::string ret;
return ret;
}
+// Dwarf_line_info routines.
+
+static unsigned int next_generation_count = 0;
+
+struct Addr2line_cache_entry
+{
+ Object* object;
+ unsigned int shndx;
+ Dwarf_line_info* dwarf_line_info;
+ unsigned int generation_count;
+ unsigned int access_count;
+
+ Addr2line_cache_entry(Object* o, unsigned int s, Dwarf_line_info* d)
+ : object(o), shndx(s), dwarf_line_info(d),
+ generation_count(next_generation_count), access_count(0)
+ {
+ if (next_generation_count < (1U << 31))
+ ++next_generation_count;
+ }
+};
+// We expect this cache to be small, so don't bother with a hashtable
+// or priority queue or anything: just use a simple vector.
+static std::vector<Addr2line_cache_entry> addr2line_cache;
+
+std::string
+Dwarf_line_info::one_addr2line(Object* object,
+ unsigned int shndx, off_t offset,
+ size_t cache_size)
+{
+ Dwarf_line_info* lineinfo = NULL;
+ std::vector<Addr2line_cache_entry>::iterator it;
+
+ // First, check the cache. If we hit, update the counts.
+ for (it = addr2line_cache.begin(); it != addr2line_cache.end(); ++it)
+ {
+ if (it->object == object && it->shndx == shndx)
+ {
+ lineinfo = it->dwarf_line_info;
+ it->generation_count = next_generation_count;
+ // We cap generation_count at 2^31 -1 to avoid overflow.
+ if (next_generation_count < (1U << 31))
+ ++next_generation_count;
+ // We cap access_count at 31 so 2^access_count doesn't overflow
+ if (it->access_count < 31)
+ ++it->access_count;
+ break;
+ }
+ }
+
+ // If we don't hit the cache, create a new object and insert into the
+ // cache.
+ if (lineinfo == NULL)
+ {
+ switch (parameters->size_and_endianness())
+ {
+#ifdef HAVE_TARGET_32_LITTLE
+ case Parameters::TARGET_32_LITTLE:
+ lineinfo = new Sized_dwarf_line_info<32, false>(object, shndx); break;
+#endif
+#ifdef HAVE_TARGET_32_BIG
+ case Parameters::TARGET_32_BIG:
+ lineinfo = new Sized_dwarf_line_info<32, true>(object, shndx); break;
+#endif
+#ifdef HAVE_TARGET_64_LITTLE
+ case Parameters::TARGET_64_LITTLE:
+ lineinfo = new Sized_dwarf_line_info<64, false>(object, shndx); break;
+#endif
+#ifdef HAVE_TARGET_64_BIG
+ case Parameters::TARGET_64_BIG:
+ lineinfo = new Sized_dwarf_line_info<64, true>(object, shndx); break;
+#endif
+ default:
+ gold_unreachable();
+ }
+ addr2line_cache.push_back(Addr2line_cache_entry(object, shndx, lineinfo));
+ }
+
+ // Now that we have our object, figure out the answer
+ std::string retval = lineinfo->addr2line(shndx, offset);
+
+ // Finally, if our cache has grown too big, delete old objects. We
+ // assume the common (probably only) case is deleting only one object.
+ // We use a pretty simple scheme to evict: function of LRU and MFU.
+ while (addr2line_cache.size() > cache_size)
+ {
+ unsigned int lowest_score = ~0U;
+ std::vector<Addr2line_cache_entry>::iterator lowest
+ = addr2line_cache.end();
+ for (it = addr2line_cache.begin(); it != addr2line_cache.end(); ++it)
+ {
+ const unsigned int score = (it->generation_count
+ + (1U << it->access_count));
+ if (score < lowest_score)
+ {
+ lowest_score = score;
+ lowest = it;
+ }
+ }
+ if (lowest != addr2line_cache.end())
+ {
+ delete lowest->dwarf_line_info;
+ addr2line_cache.erase(lowest);
+ }
+ }
+
+ return retval;
+}
+
+void
+Dwarf_line_info::clear_addr2line_cache()
+{
+ for (std::vector<Addr2line_cache_entry>::iterator it = addr2line_cache.begin();
+ it != addr2line_cache.end();
+ ++it)
+ delete it->dwarf_line_info;
+ addr2line_cache.clear();
+}
+
#ifdef HAVE_TARGET_32_LITTLE
template
-class Dwarf_line_info<32, false>;
+class Sized_dwarf_line_info<32, false>;
#endif
#ifdef HAVE_TARGET_32_BIG
template
-class Dwarf_line_info<32, true>;
+class Sized_dwarf_line_info<32, true>;
#endif
#ifdef HAVE_TARGET_64_LITTLE
template
-class Dwarf_line_info<64, false>;
+class Sized_dwarf_line_info<64, false>;
#endif
#ifdef HAVE_TARGET_64_BIG
template
-class Dwarf_line_info<64, true>;
+class Sized_dwarf_line_info<64, true>;
#endif
} // End namespace gold.