Notes on the GNU Implementation of DWARF Debugging Information
--------------------------------------------------------------
Last Updated: Sun Jul 17 08:17:42 PDT 1994 by rfg@segfault.us.com
------------------------------------------------------------

This file describes special and unique aspects of the GNU implementation
of the DWARF debugging information language, as provided in the GNU version
2.x compiler(s).

For general information about the DWARF debugging information language,
you should obtain the DWARF version 1 specification document (and perhaps
also the DWARF version 2 draft specification document) developed by the
UNIX International Programming Languages Special Interest Group.  A copy
of the DWARF version 1 specification (in PostScript form) may be
obtained either from me <rfg@netcom.com> or from the main Data General
FTP server.  (See below.)  The file you are looking at now only describes
known deviations from the DWARF version 1 specification, together with
those things which are allowed by the DWARF version 1 specification but
which are known to cause interoperability problems (e.g. with SVR4 SDB).

To obtain a copy of the DWARF Version 1 and/or DWARF Version 2 specification
from Data General's FTP server, use the following procedure:

---------------------------------------------------------------------------
	ftp to machine: "dg-rtp.dg.com" (128.222.1.2).  
 
	Log in as "ftp".
	cd to "plsig"
	get any of the following file you are interested in:
 
	dwarf.1.0.3.ps
	dwarf.2.0.0.index.ps
	dwarf.2.0.0.ps
---------------------------------------------------------------------------

The generation of DWARF debugging information by the GNU version 2.x C
compiler has now been tested rather extensively for m88k, i386, i860, and
Sparc targets.  The DWARF output of the GNU C compiler appears to inter-
operate well with the standard SVR4 SDB debugger on these kinds of target
systems (but of course, there are no guarantees).

DWARF generation for the GNU g++ compiler is still not operable.  This is
due primarily to the many remaining cases where the g++ front end does not
conform to the conventions used in the GNU C front end for representing
various kinds of declarations in the TREE data structure.  It is not clear
at this time how these problems will be addressed.

Future plans for the dwarfout.c module of the GNU compiler(s) includes the
addition of full support for GNU FORTRAN.  (This should, in theory, be a
lot simpler to add than adding support for g++... but we'll see.)

Many features of the DWARF version 2 specification have been adapted to
(and used in) the GNU implementation of DWARF (version 1).  In most of
these cases, a DWARF version 2 approach is used in place of (or in addition
to) DWARF version 1 stuff simply because it is apparent that DWARF version
1 is not sufficiently expressive to provide the kinds of information which
may be necessary to support really robust debugging.  In all of these cases
however, the use of DWARF version 2 features should not interfere in any
way with the interoperability (of GNU compilers) with generally available
"classic" (pre version 1) DWARF consumer tools (e.g. SVR4 SDB).

The DWARF generation enhancement for the GNU compiler(s) was initially
donated to the Free Software Foundation by Network Computing Devices.
(Thanks NCD!) Additional development and maintenance of dwarfout.c has
been largely supported (i.e. funded) by Intel Corporation.  (Thanks Intel!)

If you have questions or comments about the DWARF generation feature, please
send mail to me <rfg@netcom.com>.  I will be happy to investigate any bugs
reported and I may even provide fixes (but of course, I can make no promises).

The DWARF debugging information produced by GCC may deviate in a few minor
(but perhaps significant) respects from the DWARF debugging information
currently produced by other C compilers.  A serious attempt has been made
however to conform to the published specifications, to existing practice,
and to generally accepted norms in the GNU implementation of DWARF.

    ** IMPORTANT NOTE **    ** IMPORTANT NOTE **    ** IMPORTANT NOTE **

Under normal circumstances, the DWARF information generated by the GNU
compilers (in an assembly language file) is essentially impossible for
a human being to read.  This fact can make it very difficult to debug
certain DWARF-related problems.  In order to overcome this difficulty,
a feature has been added to dwarfout.c (enabled by the -fverbose-asm
option) which causes additional comments to be placed into the assembly
language output file, out to the right-hand side of most bits of DWARF
material.  The comments indicate (far more clearly that the obscure
DWARF hex codes do) what is actually being encoded in DWARF.  Thus, the
-fverbose-asm option can be highly useful for those who must study the
DWARF output from the GNU compilers in detail.

---------

(Footnote: Within this file, the term `Debugging Information Entry' will
be abbreviated as `DIE'.)


Release Notes  (aka known bugs)
-------------------------------

In one very obscure case involving dynamically sized arrays, the DWARF
"location information" for such an array may make it appear that the
array has been totally optimized out of existence, when in fact it
*must* actually exist.  (This only happens when you are using *both* -g
*and* -O.)  This is due to aggressive dead store elimination in the
compiler, and to the fact that the DECL_RTL expressions associated with
variables are not always updated to correctly reflect the effects of
GCC's aggressive dead store elimination.

-------------------------------

When attempting to set a breakpoint at the "start" of a function compiled
with -g1, the debugger currently has no way of knowing exactly where the
end of the prologue code for the function is.  Thus, for most targets,
all the debugger can do is to set the breakpoint at the AT_low_pc address
for the function.  But if you stop there and then try to look at one or
more of the formal parameter values, they may not have been "homed" yet,
so you may get inaccurate answers (or perhaps even addressing errors).

Some people may consider this simply a non-feature, but I consider it a
bug, and I hope to provide some GNU-specific attributes (on function
DIEs) which will specify the address of the end of the prologue and the
address of the beginning of the epilogue in a future release.

-------------------------------

It is believed at this time that old bugs relating to the AT_bit_offset
values for bit-fields have been fixed.

There may still be some very obscure bugs relating to the DWARF description
of type `long long' bit-fields for target machines (e.g. 80x86 machines)
where the alignment of type `long long' data objects is different from
(and less than) the size of a type `long long' data object.

Please report any problems with the DWARF description of bit-fields as you
would any other GCC bug.  (Procedures for bug reporting are given in the
GNU C compiler manual.)

--------------------------------

At this time, GCC does not know how to handle the GNU C "nested functions"
extension.  (See the GCC manual for more info on this extension to ANSI C.)

--------------------------------

The GNU compilers now represent inline functions (and inlined instances
thereof) in exactly the manner described by the current DWARF version 2
(draft) specification.  The version 1 specification for handling inline
functions (and inlined instances) was known to be brain-damaged (by the
PLSIG) when the version 1 spec was finalized, but it was simply too late
in the cycle to get it removed before the version 1 spec was formally
released to the public (by UI).

--------------------------------

At this time, GCC does not generate the kind of really precise information
about the exact declared types of entities with signed integral types which
is required by the current DWARF draft specification.

Specifically, the current DWARF draft specification seems to require that
the type of an non-unsigned integral bit-field member of a struct or union
type be represented as either a "signed" type or as a "plain" type,
depending upon the exact set of keywords that were used in the
type specification for the given bit-field member.  It was felt (by the
UI/PLSIG) that this distinction between "plain" and "signed" integral types
could have some significance (in the case of bit-fields) because ANSI C
does not constrain the signedness of a plain bit-field, whereas it does
constrain the signedness of an explicitly "signed" bit-field.  For this
reason, the current DWARF specification calls for compilers to produce
type information (for *all* integral typed entities... not just bit-fields)
which explicitly indicates the signedness of the relevant type to be
"signed" or "plain" or "unsigned".

Unfortunately, the GNU DWARF implementation is currently incapable of making
such distinctions.

--------------------------------


Known Interoperability Problems
-------------------------------

Although the GNU implementation of DWARF conforms (for the most part) with
the current UI/PLSIG DWARF version 1 specification (with many compatible
version 2 features added in as "vendor specific extensions" just for good
measure) there are a few known cases where GCC's DWARF output can cause
some confusion for "classic" (pre version 1) DWARF consumers such as the
System V Release 4 SDB debugger.  These cases are described in this section.

--------------------------------

The DWARF version 1 specification includes the fundamental type codes
FT_ext_prec_float, FT_complex, FT_dbl_prec_complex, and FT_ext_prec_complex.
Since GNU C is only a C compiler (and since C doesn't provide any "complex"
data types) the only one of these fundamental type codes which GCC ever
generates is FT_ext_prec_float.  This fundamental type code is generated
by GCC for the `long double' data type.  Unfortunately, due to an apparent
bug in the SVR4 SDB debugger, SDB can become very confused wherever any
attempt is made to print a variable, parameter, or field whose type was
given in terms of FT_ext_prec_float.

(Actually, SVR4 SDB fails to understand *any* of the four fundamental type
codes mentioned here.  This will fact will cause additional problems when
there is a GNU FORTRAN front-end.)

--------------------------------

In general, it appears that SVR4 SDB is not able to effectively ignore
fundamental type codes in the "implementation defined" range.  This can
cause problems when a program being debugged uses the `long long' data
type (or the signed or unsigned varieties thereof) because these types
are not defined by ANSI C, and thus, GCC must use its own private fundamental
type codes (from the implementation-defined range) to represent these types.

--------------------------------


General GNU DWARF extensions
----------------------------

In the current DWARF version 1 specification, no mechanism is specified by
which accurate information about executable code from include files can be
properly (and fully) described.  (The DWARF version 2 specification *does*
specify such a mechanism, but it is about 10 times more complicated than
it needs to be so I'm not terribly anxious to try to implement it right
away.)

In the GNU implementation of DWARF version 1, a fully downward-compatible
extension has been implemented which permits the GNU compilers to specify
which executable lines come from which files.  This extension places
additional information (about source file names) in GNU-specific sections
(which should be totally ignored by all non-GNU DWARF consumers) so that
this extended information can be provided (to GNU DWARF consumers) in a way
which is totally transparent (and invisible) to non-GNU DWARF consumers
(e.g. the SVR4 SDB debugger).  The additional information is placed *only*
in specialized GNU-specific sections, where it should never even be seen
by non-GNU DWARF consumers.

To understand this GNU DWARF extension, imagine that the sequence of entries
in the .lines section is broken up into several subsections.  Each contiguous
sequence of .line entries which relates to a sequence of lines (or statements)
from one particular file (either a `base' file or an `include' file) could
be called a `line entries chunk' (LEC).

For each LEC there is one entry in the .debug_srcinfo section.

Each normal entry in the .debug_srcinfo section consists of two 4-byte
words of data as follows:

	(1)	The starting address (relative to the entire .line section)
		of the first .line entry in the relevant LEC.

	(2)	The starting address (relative to the entire .debug_sfnames
		section) of a NUL terminated string representing the
		relevant filename.  (This filename name be either a
		relative or an absolute filename, depending upon how the
		given source file was located during compilation.)

Obviously, each .debug_srcinfo entry allows you to find the relevant filename,
and it also points you to the first .line entry that was generated as a result
of having compiled a given source line from the given source file.

Each subsequent .line entry should also be assumed to have been produced
as a result of compiling yet more lines from the same file.  The end of
any given LEC is easily found by looking at the first 4-byte pointer in
the *next* .debug_srcinfo entry.  That next .debug_srcinfo entry points
to a new and different LEC, so the preceding LEC (implicitly) must have
ended with the last .line section entry which occurs at the 2 1/2 words
just before the address given in the first pointer of the new .debug_srcinfo
entry.

The following picture may help to clarify this feature.  Let's assume that
`LE' stands for `.line entry'.  Also, assume that `* 'stands for a pointer.


	.line section	   .debug_srcinfo section     .debug_sfnames section
	----------------------------------------------------------------

	LE  <---------------------- *
	LE			    * -----------------> "foobar.c" <---
	LE								|
	LE								|
	LE  <---------------------- *					|
	LE			    * -----------------> "foobar.h" <|	|
	LE							     |	|
	LE							     |	|
	LE  <---------------------- *				     |	|
	LE			    * ----------------->  "inner.h"  |	|
	LE							     |	|
	LE  <---------------------- *				     |	|
	LE			    * -------------------------------	|
	LE								|
	LE								|
	LE								|
	LE								|
	LE  <---------------------- *					|
	LE			    * -----------------------------------
	LE
	LE
	LE

In effect, each entry in the .debug_srcinfo section points to *both* a
filename (in the .debug_sfnames section) and to the start of a block of
consecutive LEs (in the .line section).

Note that just like in the .line section, there are specialized first and
last entries in the .debug_srcinfo section for each object file.  These
special first and last entries for the .debug_srcinfo section are very
different from the normal .debug_srcinfo section entries.  They provide
additional information which may be helpful to a debugger when it is
interpreting the data in the .debug_srcinfo, .debug_sfnames, and .line
sections.

The first entry in the .debug_srcinfo section for each compilation unit
consists of five 4-byte words of data.  The contents of these five words
should be interpreted (by debuggers) as follows:

	(1)	The starting address (relative to the entire .line section)
		of the .line section for this compilation unit.

	(2)	The starting address (relative to the entire .debug_sfnames
		section) of the .debug_sfnames section for this compilation
		unit.

	(3)	The starting address (in the execution virtual address space)
		of the .text section for this compilation unit.

	(4)	The ending address plus one (in the execution virtual address
		space) of the .text section for this compilation unit.

	(5)	The date/time (in seconds since midnight 1/1/70) at which the
		compilation of this compilation unit occurred.  This value
		should be interpreted as an unsigned quantity because gcc
		might be configured to generate a default value of 0xffffffff
		in this field (in cases where it is desired to have object
		files created at different times from identical source files
		be byte-for-byte identical).  By default, these timestamps
		are *not* generated by dwarfout.c (so that object files
		compiled at different times will be byte-for-byte identical).
		If you wish to enable this "timestamp" feature however, you
		can simply place a #define for the symbol `DWARF_TIMESTAMPS'
		in your target configuration file and then rebuild the GNU
		compiler(s).

Note that the first string placed into the .debug_sfnames section for each
compilation unit is the name of the directory in which compilation occurred.
This string ends with a `/' (to help indicate that it is the pathname of a
directory).  Thus, the second word of each specialized initial .debug_srcinfo
entry for each compilation unit may be used as a pointer to the (string)
name of the compilation directory, and that string may in turn be used to
"absolutize" any relative pathnames which may appear later on in the
.debug_sfnames section entries for the same compilation unit.

The fifth and last word of each specialized starting entry for a compilation
unit in the .debug_srcinfo section may (depending upon your configuration)
indicate the date/time of compilation, and this may be used (by a debugger)
to determine if any of the source files which contributed code to this
compilation unit are newer than the object code for the compilation unit
itself.  If so, the debugger may wish to print an "out-of-date" warning
about the compilation unit.

The .debug_srcinfo section associated with each compilation will also have
a specialized terminating entry.  This terminating .debug_srcinfo section
entry will consist of the following two 4-byte words of data:

	(1)	The offset, measured from the start of the .line section to
		the beginning of the terminating entry for the .line section.

	(2)	A word containing the value 0xffffffff.

--------------------------------

In the current DWARF version 1 specification, no mechanism is specified by
which information about macro definitions and un-definitions may be provided
to the DWARF consumer.

The DWARF version 2 (draft) specification does specify such a mechanism.
That specification was based on the GNU ("vendor specific extension")
which provided some support for macro definitions and un-definitions,
but the "official" DWARF version 2 (draft) specification mechanism for
handling macros and the GNU implementation have diverged somewhat.  I
plan to update the GNU implementation to conform to the "official"
DWARF version 2 (draft) specification as soon as I get time to do that.

Note that in the GNU implementation, additional information about macro
definitions and un-definitions is *only* provided when the -g3 level of
debug-info production is selected.  (The default level is -g2 and the
plain old -g option is considered to be identical to -g2.)

GCC records information about macro definitions and undefinitions primarily
in a section called the .debug_macinfo section.  Normal entries in the
.debug_macinfo section consist of the following three parts:

	(1)	A special "type" byte.

	(2)	A 3-byte line-number/filename-offset field.

	(3)	A NUL terminated string.

The interpretation of the second and third parts is dependent upon the
value of the leading (type) byte.

The type byte may have one of four values depending upon the type of the
.debug_macinfo entry which follows.  The 1-byte MACINFO type codes presently
used, and their meanings are as follows:

	MACINFO_start		A base file or an include file starts here.
	MACINFO_resume		The current base or include file ends here.
	MACINFO_define          A #define directive occurs here.
	MACINFO_undef           A #undef directive occur here.

(Note that the MACINFO_... codes mentioned here are simply symbolic names
for constants which are defined in the GNU dwarf.h file.)

For MACINFO_define and MACINFO_undef entries, the second (3-byte) field
contains the number of the source line (relative to the start of the current
base source file or the current include files) when the #define or #undef
directive appears.  For a MACINFO_define entry, the following string field
contains the name of the macro which is defined, followed by its definition.
Note that the definition is always separated from the name of the macro
by at least one whitespace character.  For a MACINFO_undef entry, the
string which follows the 3-byte line number field contains just the name
of the macro which is being undef'ed.

For a MACINFO_start entry, the 3-byte field following the type byte contains
the offset, relative to the start of the .debug_sfnames section for the
current compilation unit, of a string which names the new source file which
is beginning its inclusion at this point.  Following that 3-byte field,
each MACINFO_start entry always contains a zero length NUL terminated
string.

For a MACINFO_resume entry, the 3-byte field following the type byte contains
the line number WITHIN THE INCLUDING FILE at which the inclusion of the
current file (whose inclusion ends here) was initiated.  Following that
3-byte field, each MACINFO_resume entry always contains a zero length NUL
terminated string.

Each set of .debug_macinfo entries for each compilation unit is terminated
by a special .debug_macinfo entry consisting of a 4-byte zero value followed
by a single NUL byte.

--------------------------------

In the current DWARF draft specification, no provision is made for providing
a separate level of (limited) debugging information necessary to support
tracebacks (only) through fully-debugged code (e.g. code in system libraries).

A proposal to define such a level was submitted (by me) to the UI/PLSIG.
This proposal was rejected by the UI/PLSIG for inclusion into the DWARF
version 1 specification for two reasons.  First, it was felt (by the PLSIG)
that the issues involved in supporting a "traceback only" subset of DWARF
were not well understood.  Second, and perhaps more importantly, the PLSIG
is already having enough trouble agreeing on what it means to be "conforming"
to the DWARF specification, and it was felt that trying to specify multiple
different *levels* of conformance would only complicate our discussions of
this already divisive issue.  Nonetheless, the GNU implementation of DWARF
provides an abbreviated "traceback only" level of debug-info production for
use with fully-debugged "system library" code.  This level should only be
used for fully debugged system library code, and even then, it should only
be used where there is a very strong need to conserve disk space.  This
abbreviated level of debug-info production can be used by specifying the
-g1 option on the compilation command line.

--------------------------------

As mentioned above, the GNU implementation of DWARF currently uses the DWARF
version 2 (draft) approach for inline functions (and inlined instances
thereof).  This is used in preference to the version 1 approach because
(quite simply) the version 1 approach is highly brain-damaged and probably
unworkable.

--------------------------------


GNU DWARF Representation of GNU C Extensions to ANSI C
------------------------------------------------------

The file dwarfout.c has been designed and implemented so as to provide
some reasonable DWARF representation for each and every declarative
construct which is accepted by the GNU C compiler.  Since the GNU C
compiler accepts a superset of ANSI C, this means that there are some
cases in which the DWARF information produced by GCC must take some
liberties in improvising DWARF representations for declarations which
are only valid in (extended) GNU C.

In particular, GNU C provides at least three significant extensions to
ANSI C when it comes to declarations.  These are (1) inline functions,
and (2) dynamic arrays, and (3) incomplete enum types.  (See the GCC
manual for more information on these GNU extensions to ANSI C.)  When
used, these GNU C extensions are represented (in the generated DWARF
output of GCC) in the most natural and intuitively obvious ways.

In the case of inline functions, the DWARF representation is exactly as
called for in the DWARF version 2 (draft) specification for an identical
function written in C++; i.e. we "reuse" the representation of inline
functions which has been defined for C++ to support this GNU C extension.

In the case of dynamic arrays, we use the most obvious representational
mechanism available; i.e. an array type in which the upper bound of
some dimension (usually the first and only dimension) is a variable
rather than a constant.  (See the DWARF version 1 specification for more
details.)

In the case of incomplete enum types, such types are represented simply
as TAG_enumeration_type DIEs which DO NOT contain either AT_byte_size
attributes or AT_element_list attributes.

--------------------------------


Future Directions
-----------------

The codes, formats, and other paraphernalia necessary to provide proper
support for symbolic debugging for the C++ language are still being worked
on by the UI/PLSIG.  The vast majority of the additions to DWARF which will
be needed to completely support C++ have already been hashed out and agreed
upon, but a few small issues (e.g. anonymous unions, access declarations)
are still being discussed.  Also, we in the PLSIG are still discussing
whether or not we need to do anything special for C++ templates.  (At this
time it is not yet clear whether we even need to do anything special for
these.) 

Unfortunately, as mentioned above, there are quite a few problems in the
g++ front end itself, and these are currently responsible for severely
restricting the progress which can be made on adding DWARF support
specifically for the g++ front-end.  Furthermore, Richard Stallman has
expressed the view that C++ friendships might not be important enough to
describe (in DWARF).  This view directly conflicts with both the DWARF
version 1 and version 2 (draft) specifications, so until this small
misunderstanding is cleared up, DWARF support for g++ is unlikely.

With regard to FORTRAN, the UI/PLSIG has defined what is believed to be a
complete and sufficient set of codes and rules for adequately representing
all of FORTRAN 77, and most of Fortran 90 in DWARF.  While some support for
this has been implemented in dwarfout.c, further implementation and testing
will have to await the arrival of the GNU Fortran front-end (which is
currently in early alpha test as of this writing).

GNU DWARF support for other languages (i.e. Pascal and Modula) is a moot
issue until there are GNU front-ends for these other languages.

GNU DWARF support for DWARF version 2 will probably not be attempted until
such time as the version 2 specification is finalized.  (More work needs
to be done on the version 2 specification to make the new "abbreviations"
feature of version 2 more easily implementable.  Until then, it will be
a royal pain the ass to implement version 2 "abbreviations".)  For the
time being, version 2 features will be added (in a version 1 compatible
manner) when and where these features seem necessary or extremely desirable.

As currently defined, DWARF only describes a (binary) language which can
be used to communicate symbolic debugging information from a compiler
through an assembler and a linker, to a debugger.  There is no clear
specification of what processing should be (or must be) done by the
assembler and/or the linker.  Fortunately, the role of the assembler
is easily inferred (by anyone knowledgeable about assemblers) just by
looking  at examples of assembly-level DWARF code.  Sadly though, the
allowable (or required) processing steps performed by a linker are
harder to infer and (perhaps) even harder to agree upon.  There are
several forms of very useful `post-processing' steps which intelligent
linkers *could* (in theory) perform on object files containing DWARF,
but any and all such link-time transformations are currently both disallowed
and unspecified.

In particular, possible link-time transformations of DWARF code which could
provide significant benefits include (but are not limited to):

	Commonization of duplicate DIEs obtained from multiple input
	(object) files.

	Cross-compilation type checking based upon DWARF type information
	for objects and functions.

	Other possible `compacting' transformations designed to save disk
	space and to reduce linker & debugger I/O activity.