bfd/doc/coffcode.texi

@section coff backends
BFD supports a number of different flavours of coff format.
The major differences between formats are the sizes and
alignments of fields in structures on disk, and the occasional
extra field.

Coff in all its varieties is implemented with a few common
files and a number of implementation specific files. For
example, the i386 coff format is implemented in the file
@file{coff-i386.c}.  This file @code{#include}s
@file{coff/i386.h} which defines the external structure of the
coff format for the i386, and @file{coff/internal.h} which
defines the internal structure. @file{coff-i386.c} also
defines the relocations used by the i386 coff format
@xref{Relocations}.

@subsection Porting to a new version of coff
The recommended method is to select from the existing
implementations the version of coff which is most like the one
you want to use.  For example, we'll say that i386 coff is
the one you select, and that your coff flavour is called foo.
Copy @file{i386coff.c} to @file{foocoff.c}, copy
@file{../include/coff/i386.h} to @file{../include/coff/foo.h},
and add the lines to @file{targets.c} and @file{Makefile.in}
so that your new back end is used. Alter the shapes of the
structures in @file{../include/coff/foo.h} so that they match
what you need. You will probably also have to add
@code{#ifdef}s to the code in @file{coff/internal.h} and
@file{coffcode.h} if your version of coff is too wild.

You can verify that your new BFD backend works quite simply by
building @file{objdump} from the @file{binutils} directory,
and making sure that its version of what's going on and your
host system's idea (assuming it has the pretty standard coff
dump utility, usually called @code{att-dump} or just
@code{dump}) are the same.  Then clean up your code, and send
what you've done to Cygnus. Then your stuff will be in the
next release, and you won't have to keep integrating it.

@subsection How the coff backend works


@subsubsection File layout
The Coff backend is split into generic routines that are
applicable to any Coff target and routines that are specific
to a particular target.  The target-specific routines are
further split into ones which are basically the same for all
Coff targets except that they use the external symbol format
or use different values for certain constants.

The generic routines are in @file{coffgen.c}.  These routines
work for any Coff target.  They use some hooks into the target
specific code; the hooks are in a @code{bfd_coff_backend_data}
structure, one of which exists for each target.

The essentially similar target-specific routines are in
@file{coffcode.h}.  This header file includes executable C code.
The various Coff targets first include the appropriate Coff
header file, make any special defines that are needed, and
then include @file{coffcode.h}.

Some of the Coff targets then also have additional routines in
the target source file itself.

@subsubsection Coff long section names
In the standard Coff object format, section names are limited to
the eight bytes available in the @code{s_name} field of the
@code{SCNHDR} section header structure.  The format requires the
field to be NUL-padded, but not necessarily NUL-terminated, so
the longest section names permitted are a full eight characters.

The Microsoft PE variants of the Coff object file format add
an extension to support the use of long section names.  This
extension is defined in section 4 of the Microsoft PE/COFF
specification (rev 8.1).  If a section name is too long to fit
into the section header's @code{s_name} field, it is instead
placed into the string table, and the @code{s_name} field is
filled with a slash ("/") followed by the ASCII decimal
representation of the offset of the full name relative to the
string table base.

Note that this implies that the extension can only be used in object
files, as executables do not contain a string table.  The standard
specifies that long section names from objects emitted into executable
images are to be truncated.

However, as a GNU extension, BFD can generate executable images
that contain a string table and long section names.  This
would appear to be technically valid, as the standard only says
that Coff debugging information is deprecated, not forbidden,
and in practice it works, although some tools that parse PE files
expecting the MS standard format may become confused; @file{PEview} is
one known example.

The functionality is supported in BFD by code implemented under
the control of the macro @code{COFF_LONG_SECTION_NAMES}.  If not
defined, the format does not support long section names in any way.
If defined, it is used to initialise a flag,
@code{_bfd_coff_long_section_names}, and a hook function pointer,
@code{_bfd_coff_set_long_section_names}, in the Coff backend data
structure.  The flag controls the generation of long section names
in output BFDs at runtime; if it is false, as it will be by default
when generating an executable image, long section names are truncated;
if true, the long section names extension is employed.  The hook
points to a function that allows the value of a copy of the flag
in coff object tdata to be altered at runtime, on formats that
support long section names at all; on other formats it points
to a stub that returns an error indication.

With input BFDs, the flag is set according to whether any long section
names are detected while reading the section headers.  For a completely
new BFD, the flag is set to the default for the target format.  This
information can be used by a client of the BFD library when deciding
what output format to generate, and means that a BFD that is opened
for read and subsequently converted to a writeable BFD and modified
in-place will retain whatever format it had on input.

If @code{COFF_LONG_SECTION_NAMES} is simply defined (blank), or is
defined to the value "1", then long section names are enabled by
default; if it is defined to the value zero, they are disabled by
default (but still accepted in input BFDs).  The header @file{coffcode.h}
defines a macro, @code{COFF_DEFAULT_LONG_SECTION_NAMES}, which is
used in the backends to initialise the backend data structure fields
appropriately; see the comments for further detail.

@subsubsection Bit twiddling
Each flavour of coff supported in BFD has its own header file
describing the external layout of the structures. There is also
an internal description of the coff layout, in
@file{coff/internal.h}. A major function of the
coff backend is swapping the bytes and twiddling the bits to
translate the external form of the structures into the normal
internal form. This is all performed in the
@code{bfd_swap}_@i{thing}_@i{direction} routines. Some
elements are different sizes between different versions of
coff; it is the duty of the coff version specific include file
to override the definitions of various packing routines in
@file{coffcode.h}. E.g., the size of line number entry in coff is
sometimes 16 bits, and sometimes 32 bits. @code{#define}ing
@code{PUT_LNSZ_LNNO} and @code{GET_LNSZ_LNNO} will select the
correct one. No doubt, some day someone will find a version of
coff which has a varying field size not catered to at the
moment. To port BFD, that person will have to add more @code{#defines}.
Three of the bit twiddling routines are exported to
@code{gdb}; @code{coff_swap_aux_in}, @code{coff_swap_sym_in}
and @code{coff_swap_lineno_in}. @code{GDB} reads the symbol
table on its own, but uses BFD to fix things up.  More of the
bit twiddlers are exported for @code{gas};
@code{coff_swap_aux_out}, @code{coff_swap_sym_out},
@code{coff_swap_lineno_out}, @code{coff_swap_reloc_out},
@code{coff_swap_filehdr_out}, @code{coff_swap_aouthdr_out},
@code{coff_swap_scnhdr_out}. @code{Gas} currently keeps track
of all the symbol table and reloc drudgery itself, thereby
saving the internal BFD overhead, but uses BFD to swap things
on the way out, making cross ports much safer.  Doing so also
allows BFD (and thus the linker) to use the same header files
as @code{gas}, which makes one avenue to disaster disappear.

@subsubsection Symbol reading
The simple canonical form for symbols used by BFD is not rich
enough to keep all the information available in a coff symbol
table. The back end gets around this problem by keeping the original
symbol table around, "behind the scenes".

When a symbol table is requested (through a call to
@code{bfd_canonicalize_symtab}), a request gets through to
@code{coff_get_normalized_symtab}. This reads the symbol table from
the coff file and swaps all the structures inside into the
internal form. It also fixes up all the pointers in the table
(represented in the file by offsets from the first symbol in
the table) into physical pointers to elements in the new
internal table. This involves some work since the meanings of
fields change depending upon context: a field that is a
pointer to another structure in the symbol table at one moment
may be the size in bytes of a structure at the next.  Another
pass is made over the table. All symbols which mark file names
(@code{C_FILE} symbols) are modified so that the internal
string points to the value in the auxent (the real filename)
rather than the normal text associated with the symbol
(@code{".file"}).

At this time the symbol names are moved around. Coff stores
all symbols less than nine characters long physically
within the symbol table; longer strings are kept at the end of
the file in the string table. This pass moves all strings
into memory and replaces them with pointers to the strings.

The symbol table is massaged once again, this time to create
the canonical table used by the BFD application. Each symbol
is inspected in turn, and a decision made (using the
@code{sclass} field) about the various flags to set in the
@code{asymbol}.  @xref{Symbols}. The generated canonical table
shares strings with the hidden internal symbol table.

Any linenumbers are read from the coff file too, and attached
to the symbols which own the functions the linenumbers belong to.

@subsubsection Symbol writing
Writing a symbol to a coff file which didn't come from a coff
file will lose any debugging information. The @code{asymbol}
structure remembers the BFD from which the symbol was taken, and on
output the back end makes sure that the same destination target as
source target is present.

When the symbols have come from a coff file then all the
debugging information is preserved.

Symbol tables are provided for writing to the back end in a
vector of pointers to pointers. This allows applications like
the linker to accumulate and output large symbol tables
without having to do too much byte copying.

This function runs through the provided symbol table and
patches each symbol marked as a file place holder
(@code{C_FILE}) to point to the next file place holder in the
list. It also marks each @code{offset} field in the list with
the offset from the first symbol of the current symbol.

Another function of this procedure is to turn the canonical
value form of BFD into the form used by coff. Internally, BFD
expects symbol values to be offsets from a section base; so a
symbol physically at 0x120, but in a section starting at
0x100, would have the value 0x20. Coff expects symbols to
contain their final value, so symbols have their values
changed at this point to reflect their sum with their owning
section.  This transformation uses the
@code{output_section} field of the @code{asymbol}'s
@code{asection} @xref{Sections}.

@itemize @bullet

@item
@code{coff_mangle_symbols}
@end itemize
This routine runs though the provided symbol table and uses
the offsets generated by the previous pass and the pointers
generated when the symbol table was read in to create the
structured hierarchy required by coff. It changes each pointer
to a symbol into the index into the symbol table of the asymbol.

@itemize @bullet

@item
@code{coff_write_symbols}
@end itemize
This routine runs through the symbol table and patches up the
symbols from their internal form into the coff way, calls the
bit twiddlers, and writes out the table to the file.

@findex coff_symbol_type
@subsubsection @code{coff_symbol_type}
The hidden information for an @code{asymbol} is described in a
@code{combined_entry_type}:


@example
typedef struct coff_ptr_struct
@{
  /* Remembers the offset from the first symbol in the file for
     this symbol.  Generated by coff_renumber_symbols.  */
  unsigned int offset;

  /* Selects between the elements of the union below.  */
  unsigned int is_sym : 1;

  /* Selects between the elements of the x_sym.x_tagndx union.  If set,
     p is valid and the field will be renumbered.  */
  unsigned int fix_tag : 1;

  /* Selects between the elements of the x_sym.x_fcnary.x_fcn.x_endndx
     union.  If set, p is valid and the field will be renumbered.  */
  unsigned int fix_end : 1;

  /* Selects between the elements of the x_csect.x_scnlen union.  If set,
     p is valid and the field will be renumbered.  */
  unsigned int fix_scnlen : 1;

  /* If set, u.syment.n_value contains a pointer to a symbol.  The final
     value will be the offset field.  Used for XCOFF C_BSTAT symbols.  */
  unsigned int fix_value : 1;

  /* If set, u.syment.n_value is an index into the line number entries.
     Used for XCOFF C_BINCL/C_EINCL symbols.  */
  unsigned int fix_line : 1;

  /* The container for the symbol structure as read and translated
     from the file.  */
  union
  @{
    union internal_auxent auxent;
    struct internal_syment syment;
  @} u;

 /* An extra pointer which can used by format based on COFF (like XCOFF)
    to provide extra information to their backend.  */
 void *extrap;
@} combined_entry_type;

/* Each canonical asymbol really looks like this: */

typedef struct coff_symbol_struct
@{
  /* The actual symbol which the rest of BFD works with */
  asymbol symbol;

  /* A pointer to the hidden information for this symbol */
  combined_entry_type *native;

  /* A pointer to the linenumber information for this symbol */
  struct lineno_cache_entry *lineno;

  /* Have the line numbers been relocated yet ? */
  bool done_lineno;
@} coff_symbol_type;

@end example
@findex bfd_coff_backend_data
@subsubsection @code{bfd_coff_backend_data}

@example
typedef struct
@{
  void (*_bfd_coff_swap_aux_in)
    (bfd *, void *, int, int, int, int, void *);

  void (*_bfd_coff_swap_sym_in)
    (bfd *, void *, void *);

  void (*_bfd_coff_swap_lineno_in)
    (bfd *, void *, void *);

  unsigned int (*_bfd_coff_swap_aux_out)
    (bfd *, void *, int, int, int, int, void *);

  unsigned int (*_bfd_coff_swap_sym_out)
    (bfd *, void *, void *);

  unsigned int (*_bfd_coff_swap_lineno_out)
    (bfd *, void *, void *);

  unsigned int (*_bfd_coff_swap_reloc_out)
    (bfd *, void *, void *);

  unsigned int (*_bfd_coff_swap_filehdr_out)
    (bfd *, void *, void *);

  unsigned int (*_bfd_coff_swap_aouthdr_out)
    (bfd *, void *, void *);

  unsigned int (*_bfd_coff_swap_scnhdr_out)
    (bfd *, void *, void *);

  unsigned int _bfd_filhsz;
  unsigned int _bfd_aoutsz;
  unsigned int _bfd_scnhsz;
  unsigned int _bfd_symesz;
  unsigned int _bfd_auxesz;
  unsigned int _bfd_relsz;
  unsigned int _bfd_linesz;
  unsigned int _bfd_filnmlen;
  bool _bfd_coff_long_filenames;

  bool _bfd_coff_long_section_names;
  bool (*_bfd_coff_set_long_section_names)
    (bfd *, int);

  unsigned int _bfd_coff_default_section_alignment_power;
  bool _bfd_coff_force_symnames_in_strings;
  unsigned int _bfd_coff_debug_string_prefix_length;
  unsigned int _bfd_coff_max_nscns;

  void (*_bfd_coff_swap_filehdr_in)
    (bfd *, void *, void *);

  void (*_bfd_coff_swap_aouthdr_in)
    (bfd *, void *, void *);

  void (*_bfd_coff_swap_scnhdr_in)
    (bfd *, void *, void *);

  void (*_bfd_coff_swap_reloc_in)
    (bfd *abfd, void *, void *);

  bool (*_bfd_coff_bad_format_hook)
    (bfd *, void *);

  bool (*_bfd_coff_set_arch_mach_hook)
    (bfd *, void *);

  void * (*_bfd_coff_mkobject_hook)
    (bfd *, void *, void *);

  bool (*_bfd_styp_to_sec_flags_hook)
    (bfd *, void *, const char *, asection *, flagword *);

  void (*_bfd_set_alignment_hook)
    (bfd *, asection *, void *);

  bool (*_bfd_coff_slurp_symbol_table)
    (bfd *);

  bool (*_bfd_coff_symname_in_debug)
    (bfd *, struct internal_syment *);

  bool (*_bfd_coff_pointerize_aux_hook)
    (bfd *, combined_entry_type *, combined_entry_type *,
     unsigned int, combined_entry_type *);

  bool (*_bfd_coff_print_aux)
    (bfd *, FILE *, combined_entry_type *, combined_entry_type *,
     combined_entry_type *, unsigned int);

  bool (*_bfd_coff_reloc16_extra_cases)
    (bfd *, struct bfd_link_info *, struct bfd_link_order *, arelent *,
     bfd_byte *, size_t *, size_t *);

  int (*_bfd_coff_reloc16_estimate)
    (bfd *, asection *, arelent *, unsigned int,
     struct bfd_link_info *);

  enum coff_symbol_classification (*_bfd_coff_classify_symbol)
    (bfd *, struct internal_syment *);

  bool (*_bfd_coff_compute_section_file_positions)
    (bfd *);

  bool (*_bfd_coff_start_final_link)
    (bfd *, struct bfd_link_info *);

  bool (*_bfd_coff_relocate_section)
    (bfd *, struct bfd_link_info *, bfd *, asection *, bfd_byte *,
     struct internal_reloc *, struct internal_syment *, asection **);

  reloc_howto_type *(*_bfd_coff_rtype_to_howto)
    (bfd *, asection *, struct internal_reloc *,
     struct coff_link_hash_entry *, struct internal_syment *, bfd_vma *);

  bool (*_bfd_coff_adjust_symndx)
    (bfd *, struct bfd_link_info *, bfd *, asection *,
     struct internal_reloc *, bool *);

  bool (*_bfd_coff_link_add_one_symbol)
    (struct bfd_link_info *, bfd *, const char *, flagword,
     asection *, bfd_vma, const char *, bool, bool,
     struct bfd_link_hash_entry **);

  bool (*_bfd_coff_link_output_has_begun)
    (bfd *, struct coff_final_link_info *);

  bool (*_bfd_coff_final_link_postscript)
    (bfd *, struct coff_final_link_info *);

  bool (*_bfd_coff_print_pdata)
    (bfd *, void *);

@} bfd_coff_backend_data;

@end example
@subsubsection Writing relocations
To write relocations, the back end steps though the
canonical relocation table and create an
@code{internal_reloc}. The symbol index to use is removed from
the @code{offset} field in the symbol table supplied.  The
address comes directly from the sum of the section base
address and the relocation offset; the type is dug directly
from the howto field.  Then the @code{internal_reloc} is
swapped into the shape of an @code{external_reloc} and written
out to disk.

@subsubsection Reading linenumbers
Creating the linenumber table is done by reading in the entire
coff linenumber table, and creating another table for internal use.

A coff linenumber table is structured so that each function
is marked as having a line number of 0. Each line within the
function is an offset from the first line in the function. The
base of the line number information for the table is stored in
the symbol associated with the function.

Note: The PE format uses line number 0 for a flag indicating a
new source file.

The information is copied from the external to the internal
table, and each symbol which marks a function is marked by
pointing its...

How does this work ?

@subsubsection Reading relocations
Coff relocations are easily transformed into the internal BFD form
(@code{arelent}).

Reading a coff relocation table is done in the following stages:

@itemize @bullet

@item
Read the entire coff relocation table into memory.

@item
Process each relocation in turn; first swap it from the
external to the internal form.

@item
Turn the symbol referenced in the relocation's symbol index
into a pointer into the canonical symbol table.
This table is the same as the one returned by a call to
@code{bfd_canonicalize_symtab}. The back end will call that
routine and save the result if a canonicalization hasn't been done.

@item
The reloc index is turned into a pointer to a howto
structure, in a back end specific way. For instance, the 386
uses the @code{r_type} to directly produce an index
into a howto table vector.

@item
Note that @code{arelent.addend} for COFF is often not what
most people understand as a relocation addend, but rather an
adjustment to the relocation addend stored in section contents
of relocatable object files.  The value found in section
contents may also be confusing, depending on both symbol value
and addend somewhat similar to the field value for a
final-linked object.  See @code{CALC_ADDEND}.
@end itemize