May 17, 2017
Now supports COMDAT groups as well as split-dwarf.
Some new functions added so consumer code can know
what the sections and groups contain.
April 02, 2017
The consumer code now reads DWARF5 split-dwarf sections.
The functions dwarf_init_b() and dwarf_elf_init_b()
and dwarf_object_init_b() have a groupnumber argument
added. For non-split-dwarf objects the original
dwarf_init() and dwarf_elf_init() and
dwarf_object_init() functions continue to work properly.
September 11, 2016
The producer code is getting a new batch of functions.
For each existing producer call returning
a value other than DW_DLV_OK/DW_DLV_ERROR/DW_DLV_NO_ENTRY
a new function returning one of those three codes with
a pointer argument added to return the actual desired
value (that used to be the 'return').
The existing calls require the caller to do ugly casting
to determine whether the call did its work or got an error.
The new calls have type-safe interfaces that separate
the desired result from the error-indication.
So producer calls look more like consumer calls.
The existing historical producer interfaces will remain
unchanged and will continue to be supported. The
existing interfaces will simply call the new interfaces
so there is no code duplication: Old and new use the same
implementation.
Source and binary compatibility is retained, but
users of the producer code can, when they wish,
partially or fully convert to the new interfaces.
This work will be ongoing and gradual, not 'all at once'.
August 27, 2016
Now the producer is capable of producing .debug_str strings.
Strings are never duplicated in .debug_str.
See dwarf_pro_set_default_string_form() in libdwarf2p.1.pdf
May 23, 2016
When building a shared-library the soname is set
in the dynamic section "libdwarf.so.1".
Much more testing of object correctness is done
to catch corrupted dwarf and corrupted object files.
March 14, 2016
When libdwarf callers provide neither an error argument nor
an error-handler function pointer libdwarf gives up on detecting
an error. Before now it emitted a short message on stderr
and did abort(). Now it emits a short message on stdout
and does abort(). Having libdwarf just stop application
processing (stopping the entire application) was never
good application-development policy, so one hopes no one
ever relied on doing no-error-handling themselves.
So one hopes no one will actually notice this change.
The change makes it much easier to test that the error
handling is all behaving as intended.
March 1, 2016
Much support of GNU extensions to dwarf and
of DWARF5 features now exist in libdwarf.
Including the ability to deal with the most
usual forms of split-dwarf.
Handling of zlib compression of dwarf is now
automatic (assuming you had zlib visible when
configuring libdwarf).
May 01, 2015
dwarf_next_cu_header_d() is the latest extension for CU reading.
Earlier versions still supported.
dwarf_die_from_hash_signature() added so readers can
use a hash to access a dwp package file DIE.
New support for some of DWARF5 and for dwp debug-fission
package file reading.
February 25, 2015
Now gennames uses dwgetopt() instead of getopt.
So it will more easily build where no getopt() available.
We copied dwgetopt.h, .c here (rather than reaching around
to dwarfdump) to keep the libdwarf build more easily
buildable separate from dwarfdump.
January 8, 2015
When malloc fails (out of space) dwarfdump will now
use a statically-allocated Dwarf_Error_s struct so it can get
the original malloc failure (or other error) reported back to
the calling client.
Some checks for invalid Elf files were changed to actually
report errors properly. Thanks to Edward Williamson for
providing test cases with odd errors in the Elf and DWARF.
May 19, 2014
REASON FOR RELEASE
Now handles DebugFission (part of the unreleased standard
named DWARF5) in which debug information can be split from
objects and gathered into seperate object files.
MINOR POINT:
Added a caveat that dwarf_errmsg() returned string pointers are not
necessarily pointers that can be relied on to be valid for the
duration of the current executable. It's best for consumers
to print or copy the string right away and then forget the returned
pointer. This restriction is probably already met by
most conservative consumer code: earlier, nothing was said about
the lifetime of the string! Eventually it would be nice to
actually have more ephemeral strings as then some much better
error strings could easily be produced with no leakage and
no growth in the size of the running executable.
For now, though, all dwarf_errmsg() strings continue to point to
static strings. Just please don't depend on it!
BINARY & SOURCE INCOMPATIBLITY:
The Producer code drops old dwarf_producer_init* functions
and provides an old name with a new interface: dwarf_producer_init().
The producer callback function is renamed Dwarf_Callback_Func.
Hence code calling the libdwarf producer must change.
Code calling the consumer interface is not affected.
The determination of what/how to emit DWARF is now determined
at run time, not at build time. It's now much simpler to
get the sort of output you want. Provision is made for emitting
DWARF3,4,5 and 5 (though that provision is at a very early stage
and not much supported yet).
The postal addresses for SGI in the copyright comments
changed from time to time and are no longer accurate.
So they are gradually being removed.
March 17, 2014
The dwarf.v2.mm and index.v2.mm and their pdf files have been
removed from the distribution. Those files are available
on dwarfstd.org and they are 20 years old, so removed
from libdwarf.
January 29, 2014
Now using dwarf_tsearch() instead of the original complicated
allocation code.
October 14, 2013
The Callback_Func declarations in libdwarf.h were missing a
const on the name argument. Adding it removes compiler warnings.
But removing it means client code calling
producer functions has to change to get their
callback function declarations/definitions to match.
The change does not affect those calling the consumer interfaces.
August 15, 2013
Now the printf (stdout) here go through a callback function instead
of actually using stdout. That way dwarfdump and other apps have full
control of libdwarf output from dwarf_print_lines() for example.
It also means callers of dwarf_print_lines() need to have called
dwarf_register_printf_callback() to get any actual print output.
January 26, 2013
Retracted the incompatible change. The interfaces are
again compatible with previous releases and the January 25
release is withdrawn. Has all the fixes of Jan 25 present
in the code.
January 25, 2013
The definition of the Dwarf_Loc struct
could not handle DW_OP_GNU_const_type
so it had to change. The new field added makes this version of libdwarf
incompatible with any existing dwarfdump. Rebuild dwarfdump
and dwarfdump2 to use it with this libdwarf.
Moved firmly into C89/C90 usage by more complete use of const
and be using int x(void) for example to prototype parameter-less
functions. Compiles close to cleanly with
gcc options -Wsystem-headers
-Wall -Wsign-compare
-Wno-format-y2k -W
-Wno-unused-parameter -Wstrict-prototypes
-Wmissing-prototypes -Wpointer-arith
-Wreturn-type -Wcast-qual -Wwrite-strings
-Wswitch -Wshadow -Wunused-parameter
-Wcast-align -Wchar-subscripts -Winline
-Wnested-externs -Wredundant-decls
-Wold-style-definition -Wno-pointer-sign
November 29, 2012
The function dwarf_formflag() now returns the actual flag value
instead of just 1 or 0. It has been coded wrong a very long time.
The DWARF documents (DWARF2 on) have always made it clear any non-zero
value means true and zero means false. Now consumers can properly
note the value the compiler actually put into the flag byte.
November 17, 2012
New headers contain the relocation codes for various object
ABIs. These headers are needed in libdwarf and expected
by builds of dwarfdump[2] for best dwarfdump[2] relocation
handling.
It is best to build libdwarf and dwarfdump together for
best handling of relocatable objects.
GNU compilers may generate the operator
DW_OP_GNU_implicit_pointer which is generated differently in
DWARF2 versus DWARF3/4. Hence a new interface to libdwarf
dwarf_loclist_from_expr_b() adds the compilation unit
DWARF version number to the argument list (as compared
to dwarf_loclist_from_expr_a()). Hopefully few
consumers will need to change to use the new interface.
December 13, 2011
dwarf_lineoff() is now deprecated, dwarf_lineoff_b() is
strongly recommended instead.
dwarf_add_line_entry() does not have all the line fields
needed for generating DWARF3/4, use dwarf_add_line_entry_b() instead.
Generation of DWARF3/4 is not yet functional, this new function
is a first step.
October 29, 2011
Added support for reading .debug_types (type unit) data.
October 26,2011
Revised the Makefile.in and README to make building libdwarf
easier to accomplish with unusual locations of libelf headers
or other headers or libraries.
June 04,2011
Non-Elf objects could be used with libdwarf, but
no one has contributed non-elf-reading code for libdwarf
and a crucial detail was not documented so those writing
such object-reading code have not done it entirely correctly.
Fundamentally such code must treat a section index of 0 as
a real but empty section with no name (an empty name).
dwarf_elf_access.c and dwarf_elf_init_finish.c have some comments
on this point now.
March 29,2011
All the code changed a lot because indentations were all over
the map, now they are consistent.
Additions were made to DWARF4 support.
Now we use dicheck (a new open source application) to check indentation.
Library users will not see any change, all interfaces remain
as before.
January 12,2010
A libdwarf user has noticed that the April 4, 2009
consumer function changes introduced
a problem: the default CFA column was DW_FRAME_CFA_COL
even when a newer DWARF3 consumer frame interface like
dwarf_get_fde_info_for_all_regs3() is used. The
libdwarf2.1.pdf documentation
stated the default should be DW_FRAME_CFA_COL3 in that case.
The introduction of a caller-specified frame-column
function (dwarf_set_frame_cfa_value())
in that April 4, 2009 release was flawed in that it failed
to match the documentation.
Now the default frame column is DW_FRAME_CFA_COL3 unless
the configure option --enable-oldframecol
is used at libdwarf build time.
If you are using libdwarf old frame consumer interfaces
dwarf_get_fde_info_for_reg(), dwarf_get_fde_info_for_cfa_reg(),
and dwarf_get_fde_info_for_all_regs() and want unchanged operation
then please configure libdwarf with --enable-oldframecol .
or add the call dwarf_set_frame_cfa_value(dbg,DW_FRAME_CFA_COL)
after calling a libdwarf initialization function.
It is impossible to configure a single libdwarf.a so that
it transparently defaults to both DW_FRAME_CFA_COL
and DW_FRAME_CFA_COL3.
A call such as dwarf_set_frame_cfa_value(dbg,DW_FRAME_CFA_COL3)
or dwarf_set_frame_cfa_value(dbg,DW_FRAME_CFA_COL)
(or some other name/value of your choosing)
following the dwarf_init() call gives your application
full control of the frame cfa column independent of the libdwarf
configure option. See the libdwarf2.1.pdf documentation for details.
We strongly recommend that you use dwarf_set_frame_cfa_value()
to avoid a configure-time dependency.
July 7, 2009
Implemented support for elf 'rela' relocations so libdwarf and dwarfdump
can read *nix .o files with such relocations reasonably,
at least for some machines
(see dwarf_elf_access.c for EM_ in 'case' statements.)
This changes the binary access for non-Elf object users (folks
who have coded there own non-Elf access routines do reference
internals of dwarf_opaque.h), but the new data can be left zero
and the rest of the code should work fine.
dwarf_opaque.h gathers section data in Dwarf_Section_s structs
which simplifies the code in dwarf_init_finish.c and clarifies
what fields are section related.
July 4, 2009
When something erroneous is detected in a die
information about the CU context may be of interest.
So we added dwarf_CU_dieoffset_given_die(), a function which allows
clients to find the relevant CU die for any die.
The consumer can use normal attribute access functions to
print information about that CU die (and the erroneous die,
of course). See the libdwarf consumer document for more
information.
April 27, 2009
Interface additions: dwarf_loclist_from_expr_a() and
dwarf_get_ranges_a() are new interfaces like
dwarf_loclist_from_expr() and dwarf_get_ranges() respectively,
but with arguments allowing
full support for different CIEs
in an executable having different address-sizes
(and their compilation unit DIEs if .debug_info is present).
dwarf_get_loclist_entry() does not support
differing address sizes per CIE/CU.
April 4, 2009
Added new functions dwarf_set_frame_cfa_value()
dwarf_set_frame_same_value(), and
dwarf_set_frame_undefined_value(). These are essential
for ABIs where the real register numbers exceed 1033
(such as ppc). Failing to use these leads to
frame instructions DW_CFA_undefined and
DW_CFA_same_value emitting values that cannot be
interpreted correctly by a libdwarf consumer.
See dwarfdump for examples of use.
Feb 14, 2009
Added configure option --enable-nonstandardprintf
which makes it easy to get printf of Dwarf_Unsigned (etc)
types correct even for non-standard compilers.
Dec 30, 2008
Added interfaces for getting and printing the .debug_ranges
data.
Dec 8, 2008
Record the abbreviation 'code' (index) in each DIE.
Making it possible for a pretty-printer to print the
abbreviation code.
Sep 30, 2008
Phil Mucci provided an a.out test chase which demonstrates a bug
in 64bit DWARF2 output by gcc. Now libdwarf works around this
and with -v -v -v -v prints a warning.
Sep 29, 2008
Thanks to Phil Mucci for providing a little-endian 64bit
test object file that exposed a problem when there are 'extra' bytes
(possibly unused) after a line table prologue header and before the
line table itself. This releases fixes the bug.
Thanks to Matthew Legendre for pointing out that we were sharing
de_fde_count for eh and non-eh and that could cause erroneous
error returns in a couple of functions.
These counts are now separate.
April 9, 2008
libdwarf would behave badly if one compilation unit had more than 64K
abbreviations: It was both very slow dealing with abbreviations and would
get mixed up and error-off. Increased the size of some internal variables
and rewrote abbreviation lookup.
February 18, 2008
It is now possible to write one's own access to objects, making it possible
to use a different library than libelf or even read a completely
different object format than ELF.
See dwarf_object_init() and see the new source files
dwarf_original_elf_init.c and dwarf_elf_access.c for example
code using the new function-pointer approach as it's implementation.
Thanks to Josh Fuhs for doing the design and 99% of the work to make
this happen.
February 2, 2008
Now pro_init() defaults to standard DWARF3 generated offset sizes.
But if a new flag DW_DLC_OFFSET_SIZE_64 or'd into flags passed to
dwarf_produser_init() or dwarf_producer_init_b,
the DWARF3 extended offset size is generated (if the address size
is 64 bit).
The new configure option --enable-dwarf-format-strict-32bit
forces pro_init() to always cause 32bit offset dwarf generation.
The new configure option --enable-dwarf-format-sgi-irix forces
the old SGI IRIX 64bit offset generation for 64bit pointer size objects.
This is intended to simplify standard DWARF3 generation with
the now-normal use of 32bit DWARF offsets for both 32 and 64
bit pointer objects.
It does require that anyone wanting SGI IRIX dwarf generation
with its non-standard offsets for 64bit objects use the new
--enable-dwarf-format-sgi-irix configure time option.
This has no effect on dwarf reader code. It affects code calling
the libdwarf producer interfaces.
December 8, 2007
Had to add an ugly configure conditional as libelf has
unconditional use of off64_t in recent libelf.h
July 3, 2007
A new interface function, dwarf_loclist_from_expr(),
allows easy extraction of dwarf expression bytes from
expressions in frame data.
May 8, 2007
Now documents released as .mm and .pdf (no longer as .ps).
May 7, 2007
Incorporates Sun Microsystems extensions to dwarf.h and
to the consumer and producer libraries. The changes
include corrections so the producer library cleans up it's memory
use on a call to dwarf_producer_finish(dbg).
Thanks to Chris Quenelle of Sun for these contributions.
March 20, 2007
nroff/troff and the AT&T -mm package are not widely available,
so now the Makefile refers to groff, which works quite nicely.
February 20, 2007
Documented libdwarf thread safety in README.
Fixed memory leak in dwarf macro reading code.
Removed use of static data in dwarf macro
reading code: now uses stack/heap (for
thread safety).
February 9, 2007
Maintenance of libdwarf is now outside SGI
as David Anderson has left SGI.
March 29, 2006
The March 27, 2006 version accomodates DWARF3.
Some people have been using the library without
altering dwarf.h, libdwarf.h to accomodate
large numbers of registers. This exposed a bug
(an off-by-one error) but also makes it clear
additional documentation is needed. So
in libdwarf large new comments near 'TARGET DEPENDENCY'
attempt to explain better.
Oct 03, 2005
The July version had an incompatible interface: old
dealloc code did not always work right. The incompatibility
is now fixed and the new features remain.
July 15, 2005
New optional alloc-check code optionally checks all
allocated memory is freed (malloc_check.h malloc_check.c)
Various new dealloc routines written as the previous approach
of letting client code do detailed dealloc turned out not
to dealloc all memory.
To get the new checking you must manually change a line
in malloc_check.h and rebuild libdwarf.
Mar 31, 2005
Documented the libexc.so/.debug_funcnames
dependency and the 64bit-offset DWARF extension in
mips_extentions.{mm,ps}.
Mar 21, 2005
gcc 3.3 and 3.4 .eh_frame 'z' augmentations are not handled
correctly, so libdwarf gives an error when attempting to
print such. gcc 2 'eh' augmentation is simpler and
prints correctly. (.eh_frame is a GNU section,
not DWARF2/3, and what is recorded in .eh_frame is not
specified by DWARF2/3, though .eh_frame does resemble
DWARF2/3 .debug_frame).
Oct 28, 2004
Updated contact address in copyright: SGI moved 1/4 mile
in 2003 to a new address: 1500 Crittenden Lane.
Documented additional vendor extensions.
Oct 27, 2004
Added known vendor extensions to dwarf2/3 to dwarf.h
HP, GNU, PGI and UPC extensions are now recorded.
Recorded vendor extensions from Concurrent.
Feb 3, 2004
If 'Dwarf_Word' is 64 bits, two macros reading leb numbers
fail to initialize upper bits of the values read.
First noticed with bogus line numbers printing from dwarfdump.
Now we use already-existing functions, avoiding the problem.
Oct 02, 2003
Support .debug_loc section fully.
Sept 29, 2003
Support DW_FORM_indirect properly.
Supports loclists in part (but not multiple loclist entries yet).
Support 'padding bytes' at end of .debug_arange and
.debug_pubnames and .debug_pubtypes per CU
(recent dwarf committee email made it clear this is appropriate).
May 23, 2002
Libdwarf now asks for sections only when they are
used, so that unneeded sections aren't loaded.
Support for using SGI's ELF library as an alternative to
using AT&T libelf-style has been added (the SGI ELF
library is presently only available internally to SGI).
Jan 10, 2002
Fixed memory leak in dwarf_finish().
Aug 21, 2001
If one called dwarf_add_file_decl()
or dwarf_add_directory_decl() but never added a line,
.debug_line was not produced. This was a mistake,
as if any file or directory was provided .debug_line
should be produced. Now it is produced.
June 14, 2001
Given a cu header offset, it was not easy to derive the
CU header DIE offset. Created the new
function dwarf_get_cu_die_offset_given_cu_header_offset()
do get the CU header DIE offset.
Added the function dwarf_get_arange_cu_header_offset()
so the cu header offset could be retrieved from .debug_aranges
information.
June 07, 2001
Major bug in dwarf_leb.c decoding large integers
(Dwarf_Signed 64 bit where library is compiled in ILP32)
found and fixed.
May 21, 2001
Some small fixes have been found by various folks,
so it seems time to prepare a new source release.
See ChangeLog for details.
April 15, 2000
The libdwarf copyright has changed to
version 2.1 of the GNU Lesser General Public License.
Anyone holding a version of libdwarf that was published
before this new copyright is allowed to use
the copyright published in that earlier libdwarf source
on the earlier source
or to use
this new copyright on the earlier source,
at their option.
December 08, 1999
The dwarf committee has adopted the offset-extension
proposal. This allows compatibly emitting
dwarf with 64bit offsets.
The dwarf reader now automatically figures out which is in use.
The dwarf writer configures itself at the time the
writer initialization routine is called, though
the writer is restricted, at libdwarf
compile time, to one of
mips/sgi pure 32/pure 64 offsets/pointers.
32bit offsets only (per dwarf 2.0.0 and cygnus)
32bit offsets with extension to 64bit offsets
allowed (the offset-extension newly passed).
In addition, a great deal of duplicate code
for the sgi .debug_weaknames, .debug_funcnames,
.debug_varnames and .debug_typenames sections has
been removed: a single set of functions does the real work now.
Sept 29, 1999
Just found out that cygnus is, on 64bit targets, generating
32bit offsets (as elf32 has, for example) with 64 bit
pointers (in references to text and data).
Whereas sgi has always generated 64bit dwarf with
64 bit offsets (as in elf64) and 64bit pointers for
64bit pointer objects.
I'll call the sgi approach 64-bit and the cygnus approach
32bit-offsets.
Cygnus is following the DWARF2 spec as written, so they are
right in doing only 32bit-offsets.
Folks at sgi (including me) think that, as for elf64,
the offsets in dwarf for 64bit pointer-apps should be
64 bits. We think it is only a matter of time
before we really *need* 64bit offsets and when that happens
it will be on an important app. Disk space is cheap,
so lets just go 64 bit on 64bit apps (such as ia64 apps)
to avoid a future problem.
I(davea@sgi.com) think the 'pointer-size' references in the dwarf
spec were really written for 64-bit pointer apps.
I don't recall serious consideration of 64bit pointer
apps in the committee deliberations (I did miss
a couple of meetings) and think 64bit offsets
are consistent with dwarf2, even though the speci
was not written for such. We think true full 64 bit
dwarf2 is the right way to go (the spec changes
are obvious: file and section offsets become 64bit
with 64bit pointer objects.
MIPS/SGI is definitely 64-bit offsets for 64 bit objects,
cygnus is definitely 32bit-offsets for earlier 64bit pointer
environments.
At any rate, now the dwarf reader allows and accomodates
both and the dwarf producer also accomodates both.
Some tweaking of the pro_init.c or dwarf_init_finish.c
files may be necessary in future: no other changes should
be needed to accomodate the two 64bit approaches, as
the library (and dwarfdump) now deal with both forms.
August 20, 1999
Added some #ifndef/#define to pro_util.h to let libdwarf build
on more hosts. (since those hosts don't need the producer
code, AFAIK, zero values suffice for missing #defines.)
July 21, 1999
Now reader transparently reads either-endianness data
from an either-endianness object.
Updated dwarf.h and libdwarf.h to recognize
GNU egcs dwarf extensions and to print the egcs eh_frame
section.
June 10, 1999
gnu configure version of libdwarf made available for the
first time.
Still allows only same-endian-as-host in objects.
August, 1994
libdwarf source made available for ftp on sgigate.sgi.com
/ftp/pub
June, 1994
Consumer interface changed completely, following
"Candy Machine Interfaces" chapter from
"Writing Solid Code" by Steve Maguire (Microsoft Press).
April, 1993
Initial version of libdwarf for dwarf version 2
written at sgi.