Application
Development
Michael K. Johnson
Erik W. Troan
|
|
Linux
Application Development |
Michael K. Johnson Erik W. Troan |
Stabs refers to a format for information that describes a program
to a debugger. This format was apparently invented by
Peter Kessler at
the University of California at Berkeley, for the pdx Pascal
debugger; the format has spread widely since then.
This document is one of the few published sources of documentation on stabs. It is believed to be comprehensive for stabs used by C. The lists of symbol descriptors (see section Table of Symbol Descriptors) and type descriptors (see section Table of Type Descriptors) are believed to be completely comprehensive. Stabs for COBOL-specific features and for variant records (used by Pascal and Modula-2) are poorly documented here.
Other sources of information on stabs are Dbx and Dbxtool Interfaces, 2nd edition, by Sun, 1988, and AIX Version 3.2 Files Reference, Fourth Edition, September 1992, "dbx Stabstring Grammar" in the a.out section, page 2-31. This document is believed to incorporate the information from those two sources except where it explicitly directs you to them for more information.
The GNU C compiler compiles C source in a `.c' file into assembly language in a `.s' file, which the assembler translates into a `.o' file, which the linker combines with other `.o' files and libraries to produce an executable file.
With the `-g' option, GCC puts in the `.s' file additional debugging information, which is slightly transformed by the assembler and linker, and carried through into the final executable. This debugging information describes features of the source file like line numbers, the types and scopes of variables, and function names, parameters, and scopes.
For some object file formats, the debugging information is encapsulated in assembler directives known collectively as stab (symbol table) directives, which are interspersed with the generated code. Stabs are the native format for debugging information in the a.out and XCOFF object file formats. The GNU tools can also emit stabs in the COFF and ECOFF object file formats.
The assembler adds the information from stabs to the symbol information it places by default in the symbol table and the string table of the `.o' file it is building. The linker consolidates the `.o' files into one executable file, with one symbol table and one string table. Debuggers use the symbol and string tables in the executable as a source of debugging information about the program.
There are three overall formats for stab assembler directives,
differentiated by the first word of the stab. The name of the directive
describes which combination of four possible data fields follows. It is
either .stabs (string), .stabn (number), or .stabd
(dot). IBM's XCOFF assembler uses .stabx (and some other
directives such as .file and .bi) instead of
.stabs, .stabn or .stabd.
The overall format of each class of stab is:
.stabs "string",type,other,desc,value .stabn type,other,desc,value .stabd type,other,desc .stabx "string",value,type,sdb-type
For .stabn and .stabd, there is no string (the
n_strx field is zero; see section Symbol Information in Symbol Tables). For
.stabd, the value field is implicit and has the value of
the current file location. For .stabx, the sdb-type field
is unused for stabs and can always be set to zero. The other
field is almost always unused and can be set to zero.
The number in the type field gives some basic information about which type of stab this is (or whether it is a stab, as opposed to an ordinary symbol). Each valid type number defines a different stab type; further, the stab type defines the exact interpretation of, and possible values for, any remaining string, desc, or value fields present in the stab. See section Table of Stab Types, for a list in numeric order of the valid type field values for stab directives.
For most stabs the string field holds the meat of the debugging information. The flexible nature of this field is what makes stabs extensible. For some stab types the string field contains only a name. For other stab types the contents can be a great deal more complex.
The overall format of the string field for most stab types is:
"name:symbol-descriptor type-information"
name is the name of the symbol represented by the stab; it can contain a pair of colons (see section Defining a Symbol Within Another Type). name can be omitted, which means the stab represents an unnamed object. For example, `:t10=*2' defines type 10 as a pointer to type 2, but does not give the type a name. Omitting the name field is supported by AIX dbx and GDB after about version 4.8, but not other debuggers. GCC sometimes uses a single space as the name instead of omitting the name altogether; apparently that is supported by most debuggers.
The symbol-descriptor following the `:' is an alphabetic character that tells more specifically what kind of symbol the stab represents. If the symbol-descriptor is omitted, but type information follows, then the stab represents a local variable. For a list of symbol descriptors, see section Table of Symbol Descriptors. The `c' symbol descriptor is an exception in that it is not followed by type information. See section Constants.
type-information is either a type-number, or `type-number='. A type-number alone is a type reference, referring directly to a type that has already been defined.
The `type-number=' form is a type definition, where the number represents a new type which is about to be defined. The type definition may refer to other types by number, and those type numbers may be followed by `=' and nested definitions. Also, the Lucid compiler will repeat `type-number=' more than once if it wants to define several type numbers at once.
In a type definition, if the character that follows the equals sign is non-numeric then it is a type-descriptor, and tells what kind of type is about to be defined. Any other values following the type-descriptor vary, depending on the type-descriptor. See section Table of Type Descriptors, for a list of type-descriptor values. If a number follows the `=' then the number is a type-reference. For a full description of types, section Defining Types.
There is an AIX extension for type attributes. Following the `=' are any number of type attributes. Each one starts with `@' and ends with `;'. Debuggers, including AIX's dbx and GDB 4.10, skip any type attributes they do not recognize. GDB 4.9 and other versions of dbx may not do this. Because of a conflict with C++ (see section GNU C++ Stabs), new attributes should not be defined which begin with a digit, `(', or `-'; GDB may be unable to distinguish those from the C++ type descriptor `@'. The attributes are:
aboundary
pinteger
P
ssize
S
All of this can make the string field quite long. All versions of GDB,
and some versions of dbx, can handle arbitrarily long strings. But many
versions of dbx (or assemblers or linkers, I'm not sure which)
cretinously limit the strings to about 80 characters, so compilers which
must work with such systems need to split the .stabs directive
into several .stabs directives. Each stab duplicates every field
except the string field. The string field of every stab except the last
is marked as continued with a backslash at the end (in the assembly code
this may be written as a double backslash, depending on the assembler).
Removing the backslashes and concatenating the string fields of each
stab produces the original, long string. Just to be incompatible (or so
they don't have to worry about what the assembler does with
backslashes), AIX can use `?' instead of backslash.
To get the flavor of how stabs describe source information for a C program, let's look at the simple program:
main()
{
printf("Hello world");
}
When compiled with `-g', the program above yields the following `.s' file. Line numbers have been added to make it easier to refer to parts of the `.s' file in the description of the stabs that follows.
This simple "hello world" example demonstrates several of the stab types used to describe C language source files.
1 gcc2_compiled.: 2 .stabs "/cygint/s1/users/jcm/play/",100,0,0,Ltext0 3 .stabs "hello.c",100,0,0,Ltext0 4 .text 5 Ltext0: 6 .stabs "int:t1=r1;-2147483648;2147483647;",128,0,0,0 7 .stabs "char:t2=r2;0;127;",128,0,0,0 8 .stabs "long int:t3=r1;-2147483648;2147483647;",128,0,0,0 9 .stabs "unsigned int:t4=r1;0;-1;",128,0,0,0 10 .stabs "long unsigned int:t5=r1;0;-1;",128,0,0,0 11 .stabs "short int:t6=r1;-32768;32767;",128,0,0,0 12 .stabs "long long int:t7=r1;0;-1;",128,0,0,0 13 .stabs "short unsigned int:t8=r1;0;65535;",128,0,0,0 14 .stabs "long long unsigned int:t9=r1;0;-1;",128,0,0,0 15 .stabs "signed char:t10=r1;-128;127;",128,0,0,0 16 .stabs "unsigned char:t11=r1;0;255;",128,0,0,0 17 .stabs "float:t12=r1;4;0;",128,0,0,0 18 .stabs "double:t13=r1;8;0;",128,0,0,0 19 .stabs "long double:t14=r1;8;0;",128,0,0,0 20 .stabs "void:t15=15",128,0,0,0 21 .align 4 22 LC0: 23 .ascii "Hello, world!\12\0" 24 .align 4 25 .global _main 26 .proc 1 27 _main: 28 .stabn 68,0,4,LM1 29 LM1: 30 !#PROLOGUE# 0 31 save %sp,-136,%sp 32 !#PROLOGUE# 1 33 call ___main,0 34 nop 35 .stabn 68,0,5,LM2 36 LM2: 37 LBB2: 38 sethi %hi(LC0),%o1 39 or %o1,%lo(LC0),%o0 40 call _printf,0 41 nop 42 .stabn 68,0,6,LM3 43 LM3: 44 LBE2: 45 .stabn 68,0,6,LM4 46 LM4: 47 L1: 48 ret 49 restore 50 .stabs "main:F1",36,0,0,_main 51 .stabn 192,0,0,LBB2 52 .stabn 224,0,0,LBE2
The elements of the program structure that stabs encode include the name of the main function, the names of the source and include files, the line numbers, procedure names and types, and the beginnings and ends of blocks of code.
Most languages allow the main program to have any name. The
N_MAIN stab type tells the debugger the name that is used in this
program. Only the string field is significant; it is the name of
a function which is the main program. Most C compilers do not use this
stab (they expect the debugger to assume that the name is main),
but some C compilers emit an N_MAIN stab for the main
function. I'm not sure how XCOFF handles this.
Before any other stabs occur, there must be a stab specifying the source
file. This information is contained in a symbol of stab type
N_SO; the string field contains the name of the file. The
value of the symbol is the start address of the portion of the
text section corresponding to that file.
With the Sun Solaris2 compiler, the desc field contains a source-language code.
Some compilers (for example, GCC2 and SunOS4 `/bin/cc') also
include the directory in which the source was compiled, in a second
N_SO symbol preceding the one containing the file name. This
symbol can be distinguished by the fact that it ends in a slash. Code
from the cfront C++ compiler can have additional N_SO symbols for
nonexistent source files after the N_SO for the real source file;
these are believed to contain no useful information.
For example:
.stabs "/cygint/s1/users/jcm/play/",100,0,0,Ltext0 # 100 is N_SO
.stabs "hello.c",100,0,0,Ltext0
.text
Ltext0:
Instead of N_SO symbols, XCOFF uses a .file assembler
directive which assembles to a C_FILE symbol; explaining this in
detail is outside the scope of this document.
If it is useful to indicate the end of a source file, this is done with
an N_SO symbol with an empty string for the name. The value is
the address of the end of the text section for the file. For some
systems, there is no indication of the end of a source file, and you
just need to figure it ended when you see an N_SO for a different
source file, or a symbol ending in .o (which at least some
linkers insert to mark the start of a new .o file).
There are several schemes for dealing with include files: the
traditional N_SOL approach, Sun's N_BINCL approach, and the
XCOFF C_BINCL approach (which despite the similar name has little in
common with N_BINCL).
An N_SOL symbol specifies which include file subsequent symbols
refer to. The string field is the name of the file and the value is the
text address corresponding to the end of the previous include file and
the start of this one. To specify the main source file again, use an
N_SOL symbol with the name of the main source file.
The N_BINCL approach works as follows. An N_BINCL symbol
specifies the start of an include file. In an object file, only the
string is significant; the Sun linker puts data into some of the other
fields. The end of the include file is marked by an N_EINCL
symbol (which has no string field). In an object file, there is no
significant data in the N_EINCL symbol. N_BINCL and
N_EINCL can be nested.
If the linker detects that two source files have identical stabs between
an N_BINCL and N_EINCL pair (as will generally be the case
for a header file), then it only puts out the stabs once. Each
additional occurance is replaced by an N_EXCL symbol. I believe
the Sun (SunOS4, not sure about Solaris) linker is the only one which
supports this feature.
The SunOS4 linker sets the value of a N_BINCL symbol to the total
of all the characters in the stabs strings included in the header file,
omitting the file number. The value of an N_EXCL symbol is the
same as the value of the N_BINCL symbol it replaces. I do not
know if this information is used by anything. The N_EINCL value,
and the values of the other and description fields for all three, appear
to always be zero.
For the start of an include file in XCOFF, use the `.bi' assembler
directive, which generates a C_BINCL symbol. A `.ei'
directive, which generates a C_EINCL symbol, denotes the end of
the include file. Both directives are followed by the name of the
source file in quotes, which becomes the string for the symbol.
The value of each symbol, produced automatically by the assembler
and linker, is the offset into the executable of the beginning
(inclusive, as you'd expect) or end (inclusive, as you would not expect)
of the portion of the COFF line table that corresponds to this include
file. C_BINCL and C_EINCL do not nest.
An N_SLINE symbol represents the start of a source line. The
desc field contains the line number and the value contains the code
address for the start of that source line. On most machines the address
is absolute; for stabs in sections (see section Using Stabs in Their Own Sections), it is
relative to the function in which the N_SLINE symbol occurs.
GNU documents N_DSLINE and N_BSLINE symbols for line
numbers in the data or bss segments, respectively. They are identical
to N_SLINE but are relocated differently by the linker. They
were intended to be used to describe the source location of a variable
declaration, but I believe that GCC2 actually puts the line number in
the desc field of the stab for the variable itself. GDB has been
ignoring these symbols (unless they contain a string field) since
at least GDB 3.5.
For single source lines that generate discontiguous code, such as flow of control statements, there may be more than one line number entry for the same source line. In this case there is a line number entry at the start of each code range, each with the same line number.
XCOFF does not use stabs for line numbers. Instead, it uses COFF line
numbers (which are outside the scope of this document). Standard COFF
line numbers cannot deal with include files, but in XCOFF this is fixed
with the C_BINCL method of marking include files (see section Names of Include Files).
All of the following stabs normally use the N_FUN symbol type.
However, Sun's acc compiler on SunOS4 uses N_GSYM and
N_STSYM, which means that the value of the stab for the function
is useless and the debugger must get the address of the function from
the non-stab symbols instead. On systems where non-stab symbols have
leading underscores, the stabs will lack underscores and the debugger
needs to know about the leading underscore to match up the stab and the
non-stab symbol. BSD Fortran is said to use N_FNAME with the
same restriction; the value of the symbol is not useful (I'm not sure it
really does use this, because GDB doesn't handle this and no one has
complained).
A function is represented by an `F' symbol descriptor for a global
(extern) function, and `f' for a static (local) function. For
a.out, the value of the symbol is the address of the start of the
function; it is already relocated. For stabs in ELF, the SunPRO
compiler version 2.0.1 and GCC put out an address which gets relocated
by the linker. In a future release SunPRO is planning to put out zero,
in which case the address can be found from the ELF (non-stab) symbol.
Because looking things up in the ELF symbols would probably be slow, I'm
not sure how to find which symbol of that name is the right one, and
this doesn't provide any way to deal with nested functions, it would
probably be better to make the value of the stab an address relative to
the start of the file, or just absolute. See section Having the Linker Relocate Stabs in ELF for more information on linker relocation of stabs in ELF
files. For XCOFF, the stab uses the C_FUN storage class and the
value of the stab is meaningless; the address of the function can be
found from the csect symbol (XTY_LD/XMC_PR).
The type information of the stab represents the return type of the
function; thus `foo:f5' means that foo is a function returning type
5. There is no need to try to get the line number of the start of the
function from the stab for the function; it is in the next
N_SLINE symbol.
Some compilers (such as Sun's Solaris compiler) support an extension for
specifying the types of the arguments. I suspect this extension is not
used for old (non-prototyped) function definitions in C. If the
extension is in use, the type information of the stab for the function
is followed by type information for each argument, with each argument
preceded by `;'. An argument type of 0 means that additional
arguments are being passed, whose types and number may vary (`...'
in ANSI C). GDB has tolerated this extension (parsed the syntax, if not
necessarily used the information) since at least version 4.8; I don't
know whether all versions of dbx tolerate it. The argument types given
here are not redundant with the symbols for the formal parameters
(see section Parameters); they are the types of the arguments as they are
passed, before any conversions might take place. For example, if a C
function which is declared without a prototype takes a float
argument, the value is passed as a double but then converted to a
float. Debuggers need to use the types given in the arguments
when printing values, but when calling the function they need to use the
types given in the symbol defining the function.
If the return type and types of arguments of a function which is defined
in another source file are specified (i.e., a function prototype in ANSI
C), traditionally compilers emit no stab; the only way for the debugger
to find the information is if the source file where the function is
defined was also compiled with debugging symbols. As an extension the
Solaris compiler uses symbol descriptor `P' followed by the return
type of the function, followed by the arguments, each preceded by
`;', as in a stab with symbol descriptor `f' or `F'.
This use of symbol descriptor `P' can be distinguished from its use
for register parameters (see section Passing Parameters in Registers) by the fact that it has
symbol type N_FUN.
The AIX documentation also defines symbol descriptor `J' as an internal function. I assume this means a function nested within another function. It also says symbol descriptor `m' is a module in Modula-2 or extended Pascal.
Procedures (functions which do not return values) are represented as
functions returning the void type in C. I don't see why this couldn't
be used for all languages (inventing a void type for this purpose if
necessary), but the AIX documentation defines `I', `P', and
`Q' for internal, global, and static procedures, respectively.
These symbol descriptors are unusual in that they are not followed by
type information.
The following example shows a stab for a function main which
returns type number 1. The _main specified for the value
is a reference to an assembler label which is used to fill in the start
address of the function.
.stabs "main:F1",36,0,0,_main # 36 is N_FUN
The stab representing a procedure is located immediately following the code of the procedure. This stab is in turn directly followed by a group of other stabs describing elements of the procedure. These other stabs describe the procedure's parameters, its block local variables, and its block structure.
For any of the symbol descriptors representing procedures, after the symbol descriptor and the type information is optionally a scope specifier. This consists of a comma, the name of the procedure, another comma, and the name of the enclosing procedure. The first name is local to the scope specified, and seems to be redundant with the name of the symbol (before the `:'). This feature is used by GCC, and presumably Pascal, Modula-2, etc., compilers, for nested functions.
If procedures are nested more than one level deep, only the immediately containing scope is specified. For example, this code:
int
foo (int x)
{
int bar (int y)
{
int baz (int z)
{
return x + y + z;
}
return baz (x + 2 * y);
}
return x + bar (3 * x);
}
produces the stabs:
.stabs "baz:f1,baz,bar",36,0,0,_baz.15 # 36 is N_FUN .stabs "bar:f1,bar,foo",36,0,0,_bar.12 .stabs "foo:F1",36,0,0,_foo
The program's block structure is represented by the N_LBRAC (left
brace) and the N_RBRAC (right brace) stab types. The variables
defined inside a block precede the N_LBRAC symbol for most
compilers, including GCC. Other compilers, such as the Convex, Acorn
RISC machine, and Sun acc compilers, put the variables after the
N_LBRAC symbol. The values of the N_LBRAC and
N_RBRAC symbols are the start and end addresses of the code of
the block, respectively. For most machines, they are relative to the
starting address of this source file. For the Gould NP1, they are
absolute. For stabs in sections (see section Using Stabs in Their Own Sections), they are
relative to the function in which they occur.
The N_LBRAC and N_RBRAC stabs that describe the block
scope of a procedure are located after the N_FUN stab that
represents the procedure itself.
Sun documents the desc field of N_LBRAC and
N_RBRAC symbols as containing the nesting level of the block.
However, dbx seems to not care, and GCC always sets desc to
zero.
For XCOFF, block scope is indicated with C_BLOCK symbols. If the
name of the symbol is `.bb', then it is the beginning of the block;
if the name of the symbol is `.be'; it is the end of the block.
Some languages, like Fortran, have the ability to enter procedures at
some place other than the beginning. One can declare an alternate entry
point. The N_ENTRY stab is for this; however, the Sun FORTRAN
compiler doesn't use it. According to AIX documentation, only the name
of a C_ENTRY stab is significant; the address of the alternate
entry point comes from the corresponding external symbol. A previous
revision of this document said that the value of an N_ENTRY stab
was the address of the alternate entry point, but I don't know the
source for that information.
The `c' symbol descriptor indicates that this stab represents a constant. This symbol descriptor is an exception to the general rule that symbol descriptors are followed by type information. Instead, it is followed by `=' and one of the following:
b value
c value
e type-information , value
int, but it does not
do anything violent, and future debuggers could be extended to accept
integers of any size (whether unsigned or not). This constant type is
usually documented as being only for enumeration constants, but GDB has
never imposed that restriction; I don't know about other debuggers.
i value
int); to specify
the type explicitly, use `e' instead.
r value
atof.
s string
S type-information , elements , bits , pattern
The boolean, character, string, and set constants are not supported by GDB 4.9, but it ignores them. GDB 4.8 and earlier gave an error message and refused to read symbols from the file containing the constants.
The above information is followed by `;'.
Different types of stabs describe the various ways that variables can be allocated: on the stack, globally, in registers, in common blocks, statically, or as arguments to a function.
If a variable's scope is local to a function and its lifetime is only as long as that function executes (C calls such variables automatic), it can be allocated in a register (see section Register Variables) or on the stack.
Each variable allocated on the stack has a stab with the symbol
descriptor omitted. Since type information should begin with a digit,
`-', or `(', only those characters precluded from being used
for symbol descriptors. However, the Acorn RISC machine (ARM) is said
to get this wrong: it puts out a mere type definition here, without the
preceding `type-number='. This is a bad idea; there is no
guarantee that type descriptors are distinct from symbol descriptors.
Stabs for stack variables use the N_LSYM stab type, or
C_LSYM for XCOFF.
The value of the stab is the offset of the variable within the local variables. On most machines this is an offset from the frame pointer and is negative. The location of the stab specifies which block it is defined in; see section Block Structure.
For example, the following C code:
int
main ()
{
int x;
}
produces the following stabs:
.stabs "main:F1",36,0,0,_main # 36 is N_FUN .stabs "x:1",128,0,0,-12 # 128 is N_LSYM .stabn 192,0,0,LBB2 # 192 is N_LBRAC .stabn 224,0,0,LBE2 # 224 is N_RBRAC
See section Procedures for more information on the N_FUN stab, and
section Block Structure for more information on the N_LBRAC and
N_RBRAC stabs.
A variable whose scope is not specific to just one source file is
represented by the `G' symbol descriptor. These stabs use the
N_GSYM stab type (C_GSYM for XCOFF). The type information for
the stab (see section The String Field) gives the type of the variable.
For example, the following source code:
char g_foo = 'c';
yields the following assembly code:
.stabs "g_foo:G2",32,0,0,0 # 32 is N_GSYM
.global _g_foo
.data
_g_foo:
.byte 99
The address of the variable represented by the N_GSYM is not
contained in the N_GSYM stab. The debugger gets this information
from the external symbol for the global variable. In the example above,
the .global _g_foo and _g_foo: lines tell the assembler to
produce an external symbol.
Some compilers, like GCC, output N_GSYM stabs only once, where
the variable is defined. Other compilers, like SunOS4 /bin/cc, output a
N_GSYM stab for each compilation unit which references the
variable.
Register variables have their own stab type, N_RSYM
(C_RSYM for XCOFF), and their own symbol descriptor, `r'.
The stab's value is the number of the register where the variable data
will be stored.
AIX defines a separate symbol descriptor `d' for floating point registers. This seems unnecessary; why not just just give floating point registers different register numbers? I have not verified whether the compiler actually uses `d'.
If the register is explicitly allocated to a global variable, but not initialized, as in:
register int g_bar asm ("%g5");
then the stab may be emitted at the end of the object file, with the other bss symbols.
A common block is a statically allocated section of memory which can be referred to by several source files. It may contain several variables. I believe Fortran is the only language with this feature.
A N_BCOMM stab begins a common block and an N_ECOMM stab
ends it. The only field that is significant in these two stabs is the
string, which names a normal (non-debugging) symbol that gives the
address of the common block. According to IBM documentation, only the
N_BCOMM has the name of the common block (even though their
compiler actually puts it both places).
The stabs for the members of the common block are between the
N_BCOMM and the N_ECOMM; the value of each stab is the
offset within the common block of that variable. IBM uses the
C_ECOML stab type, and there is a corresponding N_ECOML
stab type, but Sun's Fortran compiler uses N_GSYM instead. The
variables within a common block use the `V' symbol descriptor (I
believe this is true of all Fortran variables). Other stabs (at least
type declarations using C_DECL) can also be between the
N_BCOMM and the N_ECOMM.
Initialized static variables are represented by the `S' and `V' symbol descriptors. `S' means file scope static, and `V' means procedure scope static. One exception: in XCOFF, IBM's xlc compiler always uses `V', and whether it is file scope or not is distinguished by whether the stab is located within a function.
In a.out files, N_STSYM means the data section, N_FUN
means the text section, and N_LCSYM means the bss section. For
those systems with a read-only data section separate from the text
section (Solaris), N_ROSYM means the read-only data section.
For example, the source lines:
static const int var_const = 5; static int var_init = 2; static int var_noinit;
yield the following stabs:
.stabs "var_const:S1",36,0,0,_var_const # 36 is N_FUN ... .stabs "var_init:S1",38,0,0,_var_init # 38 is N_STSYM ... .stabs "var_noinit:S1",40,0,0,_var_noinit # 40 is N_LCSYM
In XCOFF files, the stab type need not indicate the section;
C_STSYM can be used for all statics. Also, each static variable
is enclosed in a static block. A C_BSTAT (emitted with a
`.bs' assembler directive) symbol begins the static block; its
value is the symbol number of the csect symbol whose value is the
address of the static block, its section is the section of the variables
in that static block, and its name is `.bs'. A C_ESTAT
(emitted with a `.es' assembler directive) symbol ends the static
block; its name is `.es' and its value and section are ignored.
In ECOFF files, the storage class is used to specify the section, so the stab type need not indicate the section.
In ELF files, for the SunPRO compiler version 2.0.1, symbol descriptor `S' means that the address is absolute (the linker relocates it) and symbol descriptor `V' means that the address is relative to the start of the relevant section for that compilation unit. SunPRO has plans to have the linker stop relocating stabs; I suspect that their the debugger gets the address from the corresponding ELF (not stab) symbol. I'm not sure how to find which symbol of that name is the right one. The clean way to do all this would be to have a the value of a symbol descriptor `S' symbol be an offset relative to the start of the file, just like everything else, but that introduces obvious compatibility problems. For more information on linker stab relocation, See section Having the Linker Relocate Stabs in ELF.
Fortran (at least, the Sun and SGI dialects of FORTRAN-77) has a feature
which allows allocating arrays with malloc, but which avoids
blurring the line between arrays and pointers the way that C does. In
stabs such a variable uses the `b' symbol descriptor.
For example, the Fortran declarations
real foo, foo10(10), foo10_5(10,5) pointer (foop, foo) pointer (foo10p, foo10) pointer (foo105p, foo10_5)
produce the stabs
foo:b6 foo10:bar3;1;10;6 foo10_5:bar3;1;5;ar3;1;10;6
In this example, real is type 6 and type 3 is an integral type
which is the type of the subscripts of the array (probably
integer).
The `b' symbol descriptor is like `V' in that it denotes a
statically allocated symbol whose scope is local to a function; see
See section Static Variables. The value of the symbol, instead of being the address
of the variable itself, is the address of a pointer to that variable.
So in the above example, the value of the foo stab is the address
of a pointer to a real, the value of the foo10 stab is the
address of a pointer to a 10-element array of reals, and the value of
the foo10_5 stab is the address of a pointer to a 5-element array
of 10-element arrays of reals.
Formal parameters to a function are represented by a stab (or sometimes two; see below) for each parameter. The stabs are in the order in which the debugger should print the parameters (i.e., the order in which the parameters are declared in the source file). The exact form of the stab depends on how the parameter is being passed.
Parameters passed on the stack use the symbol descriptor `p' and
the N_PSYM symbol type (or C_PSYM for XCOFF). The value
of the symbol is an offset used to locate the parameter on the stack;
its exact meaning is machine-dependent, but on most machines it is an
offset from the frame pointer.
As a simple example, the code:
main (argc, argv)
int argc;
char **argv;
produces the stabs:
.stabs "main:F1",36,0,0,_main # 36 is N_FUN .stabs "argc:p1",160,0,0,68 # 160 is N_PSYM .stabs "argv:p20=*21=*2",160,0,0,72
The type definition of argv is interesting because it contains
several type definitions. Type 21 is pointer to type 2 (char) and
argv (type 20) is pointer to type 21.
The following symbol descriptors are also said to go with N_PSYM.
The value of the symbol is said to be an offset from the argument
pointer (I'm not sure whether this is true or not).
pP (<<??>>) pF Fortran function parameter X (function result variable)
If the parameter is passed in a register, then traditionally there are two symbols for each argument:
.stabs "arg:p1" . . . ; N_PSYM .stabs "arg:r1" . . . ; N_RSYM
Debuggers use the second one to find the value, and the first one to know that it is an argument.
Because that approach is kind of ugly, some compilers use symbol
descriptor `P' or `R' to indicate an argument which is in a
register. Symbol type C_RPSYM is used in XCOFF and N_RSYM
is used otherwise. The symbol's value is the register number. `P'
and `R' mean the same thing; the difference is that `P' is a
GNU invention and `R' is an IBM (XCOFF) invention. As of version
4.9, GDB should handle either one.
There is at least one case where GCC uses a `p' and `r' pair rather than `P'; this is where the argument is passed in the argument list and then loaded into a register.
According to the AIX documentation, symbol descriptor `D' is for a parameter passed in a floating point register. This seems unnecessary--why not just use `R' with a register number which indicates that it's a floating point register? I haven't verified whether the system actually does what the documentation indicates.
On the sparc and hppa, for a `P' symbol whose type is a structure
or union, the register contains the address of the structure. On the
sparc, this is also true of a `p' and `r' pair (using Sun
cc) or a `p' symbol. However, if a (small) structure is
really in a register, `r' is used. And, to top it all off, on the
hppa it might be a structure which was passed on the stack and loaded
into a register and for which there is a `p' and `r' pair! I
believe that symbol descriptor `i' is supposed to deal with this
case (it is said to mean "value parameter by reference, indirect
access"; I don't know the source for this information), but I don't know
details or what compilers or debuggers use it, if any (not GDB or GCC).
It is not clear to me whether this case needs to be dealt with
differently than parameters passed by reference (see section Passing Parameters by Reference).
There is a case similar to an argument in a register, which is an argument that is actually stored as a local variable. Sometimes this happens when the argument was passed in a register and then the compiler stores it as a local variable. If possible, the compiler should claim that it's in a register, but this isn't always done.
If a parameter is passed as one type and converted to a smaller type by
the prologue (for example, the parameter is declared as a float,
but the calling conventions specify that it is passed as a
double), then GCC2 (sometimes) uses a pair of symbols. The first
symbol uses symbol descriptor `p' and the type which is passed.
The second symbol has the type and location which the parameter actually
has after the prologue. For example, suppose the following C code
appears with no prototypes involved:
void
subr (f)
float f;
{
if f is passed as a double at stack offset 8, and the prologue
converts it to a float in register number 0, then the stabs look like:
.stabs "f:p13",160,0,3,8 # 160 isN_PSYM, here 13 isdouble.stabs "f:r12",64,0,3,0 # 64 isN_RSYM, here 12 isfloat
In both stabs 3 is the line number where f is declared
(see section Line Numbers).
GCC, at least on the 960, has another solution to the same problem. It
uses a single `p' symbol descriptor for an argument which is stored
as a local variable but uses N_LSYM instead of N_PSYM. In
this case, the value of the symbol is an offset relative to the local
variables for that function, not relative to the arguments; on some
machines those are the same thing, but not on all.
On the VAX or on other machines in which the calling convention includes the number of words of arguments actually passed, the debugger (GDB at least) uses the parameter symbols to keep track of whether it needs to print nameless arguments in addition to the formal parameters which it has printed because each one has a stab. For example, in
extern int fprintf (FILE *stream, char *format, ...); ... fprintf (stdout, "%d\n", x);
there are stabs for stream and format. On most machines,
the debugger can only print those two arguments (because it has no way
of knowing that additional arguments were passed), but on the VAX or
other machines with a calling convention which indicates the number of
words of arguments, the debugger can print all three arguments. To do
so, the parameter symbol (symbol descriptor `p') (not necessarily
`r' or symbol descriptor omitted symbols) needs to contain the
actual type as passed (for example, double not float if it
is passed as a double and converted to a float).
If the parameter is passed by reference (e.g., Pascal VAR
parameters), then the symbol descriptor is `v' if it is in the
argument list, or `a' if it in a register. Other than the fact
that these contain the address of the parameter rather than the
parameter itself, they are identical to `p' and `R',
respectively. I believe `a' is an AIX invention; `v' is
supported by all stabs-using systems as far as I know.
Conformant arrays are a feature of Modula-2, and perhaps other languages, in which the size of an array parameter is not known to the called function until run-time. Such parameters have two stabs: a `x' for the array itself, and a `C', which represents the size of the array. The value of the `x' stab is the offset in the argument list where the address of the array is stored (it this right? it is a guess); the value of the `C' stab is the offset in the argument list where the size of the array (in elements? in bytes?) is stored.
The examples so far have described types as references to previously defined types, or defined in terms of subranges of or pointers to previously defined types. This chapter describes the other type descriptors that may follow the `=' in a type definition.
Certain types are built in (int, short, void,
float, etc.); the debugger recognizes these types and knows how
to handle them. Thus, don't be surprised if some of the following ways
of specifying builtin types do not specify everything that a debugger
would need to know about the type--in some cases they merely specify
enough information to distinguish the type from other types.
The traditional way to define builtin types is convolunted, so new ways
have been invented to describe them. Sun's acc uses special
builtin type descriptors (`b' and `R'), and IBM uses negative
type numbers. GDB accepts all three ways, as of version 4.8; dbx just
accepts the traditional builtin types and perhaps one of the other two
formats. The following sections describe each of these formats.
This is the traditional, convoluted method for defining builtin types.
There are several classes of such type definitions: integer, floating
point, and void.
Often types are defined as subranges of themselves. If the bounding values
fit within an int, then they are given normally. For example:
.stabs "int:t1=r1;-2147483648;2147483647;",128,0,0,0 # 128 is N_LSYM .stabs "char:t2=r2;0;127;",128,0,0,0
Builtin types can also be described as subranges of int:
.stabs "unsigned short:t6=r1;0;65535;",128,0,0,0
If the lower bound of a subrange is 0 and the upper bound is -1,
the type is an unsigned integral type whose bounds are too
big to describe in an int. Traditionally this is only used for
unsigned int and unsigned long:
.stabs "unsigned int:t4=r1;0;-1;",128,0,0,0
For larger types, GCC 2.4.5 puts out bounds in octal, with one or more leading zeroes. In this case a negative bound consists of a number which is a 1 bit (for the sign bit) followed by a 0 bit for each bit in the number (except the sign bit), and a positive bound is one which is a 1 bit for each bit in the number (except possibly the sign bit). All known versions of dbx and GDB version 4 accept this (at least in the sense of not refusing to process the file), but GDB 3.5 refuses to read the whole file containing such symbols. So GCC 2.3.3 did not output the proper size for these types. As an example of octal bounds, the string fields of the stabs for 64 bit integer types look like:
long int:t3=r1;001000000000000000000000;000777777777777777777777; long unsigned int:t5=r1;000000000000000000000000;001777777777777777777777;
If the lower bound of a subrange is 0 and the upper bound is negative,
the type is an unsigned integral type whose size in bytes is the
absolute value of the upper bound. I believe this is a Convex
convention for unsigned long long.
If the lower bound of a subrange is negative and the upper bound is 0,
the type is a signed integral type whose size in bytes is
the absolute value of the lower bound. I believe this is a Convex
convention for long long. To distinguish this from a legitimate
subrange, the type should be a subrange of itself. I'm not sure whether
this is the case for Convex.
If the upper bound of a subrange is 0 and the lower bound is positive, the type is a floating point type, and the lower bound of the subrange indicates the number of bytes in the type:
.stabs "float:t12=r1;4;0;",128,0,0,0 .stabs "double:t13=r1;8;0;",128,0,0,0
However, GCC writes long double the same way it writes
double, so there is no way to distinguish.
.stabs "long double:t14=r1;8;0;",128,0,0,0
Complex types are defined the same way as floating-point types; there is no way to distinguish a single-precision complex from a double-precision floating-point type.
The C void type is defined as itself:
.stabs "void:t15=15",128,0,0,0
I'm not sure how a boolean type is represented.
This is the method used by Sun's acc for defining builtin types.
These are the type descriptors to define builtin types:
b signed char-flag width ; offset ; nbits ;
wchar_t so the debugger
can print such variables differently (Solaris does not do this). Sun
sets it on the C types signed char and unsigned char which
arguably is wrong. width and offset appear to be for small
objects stored in larger ones, for example a short in an
int register. width is normally the number of bytes in the
type. offset seems to always be zero. nbits is the number
of bits in the type.
Note that type descriptor `b' used for builtin types conflicts with
its use for Pascal space types (see section Miscellaneous Types); they can
be distinguished because the character following the type descriptor
will be a digit, `(', or `-' for a Pascal space type, or
`u' or `s' for a builtin type.
w
R fp-type ; bytes ;
1 (NF_SINGLE)
2 (NF_DOUBLE)
3 (NF_COMPLEX)
4 (NF_COMPLEX16)
5 (NF_COMPLEX32)
complex, double complex, and
complex*16, respectively, but what does that mean? (i.e., Single
precision? Double precison?).
6 (NF_LDOUBLE)
long double, and new codes should be used for other floating
point formats (NF_DOUBLE can be used if a long double is
really just an IEEE double, of course).
g type-information ; nbits
c type-information ; nbits
The C void type is defined as a signed integral type 0 bits long:
.stabs "void:t19=bs0;0;0",128,0,0,0
The Solaris compiler seems to omit the trailing semicolon in this case. Getting sloppy in this way is not a swift move because if a type is embedded in a more complex expression it is necessary to be able to tell where it ends.
I'm not sure how a boolean type is represented.
This is the method used in XCOFF for defining builtin types. Since the debugger knows about the builtin types anyway, the idea of negative type numbers is simply to give a special type number which indicates the builtin type. There is no stab defining these types.
There are several subtle issues with negative type numbers.
One is the size of the type. A builtin type (for example the C types
int or long) might have different sizes depending on
compiler options, the target architecture, the ABI, etc. This issue
doesn't come up for IBM tools since (so far) they just target the
RS/6000; the sizes indicated below for each size are what the IBM
RS/6000 tools use. To deal with differing sizes, either define separate
negative type numbers for each size (which works but requires changing
the debugger, and, unless you get both AIX dbx and GDB to accept the
change, introduces an incompatibility), or use a type attribute
(see section The String Field) to define a new type with the appropriate size
(which merely requires a debugger which understands type attributes,
like AIX dbx or GDB). For example,
.stabs "boolean:t10=@s8;-16",128,0,0,0
defines an 8-bit boolean type, and
.stabs "boolean:t10=@s64;-16",128,0,0,0
defines a 64-bit boolean type.
A similar issue is the format of the type. This comes up most often for floating-point types, which could have various formats (particularly extended doubles, which vary quite a bit even among IEEE systems). Again, it is best to define a new negative type number for each different format; changing the format based on the target system has various problems. One such problem is that the Alpha has both VAX and IEEE floating types. One can easily imagine one library using the VAX types and another library in the same executable using the IEEE types. Another example is that the interpretation of whether a boolean is true or false can be based on the least significant bit, most significant bit, whether it is zero, etc., and different compilers (or different options to the same compiler) might provide different kinds of boolean.
The last major issue is the names of the types. The name of a given
type depends only on the negative type number given; these do not
vary depending on the language, the target system, or anything else.
One can always define separate type numbers--in the following list you
will see for example separate int and integer*4 types
which are identical except for the name. But compatibility can be
maintained by not inventing new negative type numbers and instead just
defining a new type with a new name. For example:
.stabs "CARDINAL:t10=-8",128,0,0,0
Here is the list of negative type numbers. The phrase integral type is used to mean twos-complement (I strongly suspect that all machines which use stabs use twos-complement; most machines use twos-complement these days).
-1
int, 32 bit signed integral type.
-2
char, 8 bit type holding a character. Both GDB and dbx on AIX
treat this as signed. GCC uses this type whether char is signed
or not, which seems like a bad idea. The AIX compiler (xlc) seems to
avoid this type; it uses -5 instead for char.
-3
short, 16 bit signed integral type.
-4
long, 32 bit signed integral type.
-5
unsigned char, 8 bit unsigned integral type.
-6
signed char, 8 bit signed integral type.
-7
unsigned short, 16 bit unsigned integral type.
-8
unsigned int, 32 bit unsigned integral type.
-9
unsigned, 32 bit unsigned integral type.
-10
unsigned long, 32 bit unsigned integral type.
-11
void, type indicating the lack of a value.
-12
float, IEEE single precision.
-13
double, IEEE double precision.
-14
long double, IEEE double precision. The compiler claims the size
will increase in a future release, and for binary compatibility you have
to avoid using long double. I hope when they increase it they
use a new negative type number.
-15
integer. 32 bit signed integral type.
-16
boolean. 32 bit type. GDB and GCC assume that zero is false,
one is true, and other values have unspecified meaning. I hope this
agrees with how the IBM tools use the type.
-17
short real. IEEE single precision.
-18
real. IEEE double precision.
-19
stringptr. See section Strings.
-20
character, 8 bit unsigned character type.
-21
logical*1, 8 bit type. This Fortran type has a split
personality in that it is used for boolean variables, but can also be
used for unsigned integers. 0 is false, 1 is true, and other values are
non-boolean.
-22
logical*2, 16 bit type. This Fortran type has a split
personality in that it is used for boolean variables, but can also be
used for unsigned integers. 0 is false, 1 is true, and other values are
non-boolean.
-23
logical*4, 32 bit type. This Fortran type has a split
personality in that it is used for boolean variables, but can also be
used for unsigned integers. 0 is false, 1 is true, and other values are
non-boolean.
-24
logical, 32 bit type. This Fortran type has a split
personality in that it is used for boolean variables, but can also be
used for unsigned integers. 0 is false, 1 is true, and other values are
non-boolean.
-25
complex. A complex type consisting of two IEEE single-precision
floating point values.
-26
complex. A complex type consisting of two IEEE double-precision
floating point values.
-27
integer*1, 8 bit signed integral type.
-28
integer*2, 16 bit signed integral type.
-29
integer*4, 32 bit signed integral type.
-30
wchar. Wide character, 16 bits wide, unsigned (what format?
Unicode?).
-31
long long, 64 bit signed integral type.
-32
unsigned long long, 64 bit unsigned integral type.
-33
logical*8, 64 bit unsigned integral type.
-34
integer*8, 64 bit signed integral type.
b type-information ; bytes
B type-information
d type-information
k type-information
M type-information ; length
character*3 is
represented by `M-2;3', where `-2' is a reference to a
character type (see section Negative Type Numbers). I'm not sure how this
differs from an array. This appears to be a Fortran feature.
length is a bound, like those in range types; see section Subrange Types.
S type-information
* type-information
A type can be used before it is defined; one common way to deal with that situation is just to use a type reference to a type which has not yet been defined.
Another way is with the `x' type descriptor, which is followed by `s' for a structure tag, `u' for a union tag, or `e' for a enumerator tag, followed by the name of the tag, followed by `:'. If the name contains `::' between a `<' and `>' pair (for C++ templates), such a `::' does not end the name--only a single `:' ends the name; see section Defining a Symbol Within Another Type.
For example, the following C declarations:
struct foo; struct foo *bar;
produce:
.stabs "bar:G16=*17=xsfoo:",32,0,0,0
Not all debuggers support the `x' type descriptor, so on some machines GCC does not use it. I believe that for the above example it would just emit a reference to type 17 and never define it, but I haven't verified that.
Modula-2 imported types, at least on AIX, use the `i' type descriptor, which is followed by the name of the module from which the type is imported, followed by `:', followed by the name of the type. There is then optionally a comma followed by type information for the type. This differs from merely naming the type (see section Giving a Type a Name) in that it identifies the module; I don't understand whether the name of the type given here is always just the same as the name we are giving it, or whether this type descriptor is used with a nameless stab (see section The String Field), or what. The symbol ends with `;'.
The `r' type descriptor defines a type as a subrange of another type. It is followed by type information for the type of which it is a subrange, a semicolon, an integral lower bound, a semicolon, an integral upper bound, and a semicolon. The AIX documentation does not specify the trailing semicolon, in an effort to specify array indexes more cleanly, but a subrange which is not an array index has always included a trailing semicolon (see section Array Types).
Instead of an integer, either bound can be one of the following:
A offset
T offset
a register-number
t register-number
J
Subranges are also used for builtin types; see section Traditional Builtin Types.
Arrays use the `a' type descriptor. Following the type descriptor is the type of the index and the type of the array elements. If the index type is a range type, it ends in a semicolon; otherwise (for example, if it is a type reference), there does not appear to be any way to tell where the types are separated. In an effort to clean up this mess, IBM documents the two types as being separated by a semicolon, and a range type as not ending in a semicolon (but this is not right for range types which are not array indexes, see section Subrange Types). I think probably the best solution is to specify that a semicolon ends a range type, and that the index type and element type of an array are separated by a semicolon, but that if the index type is a range type, the extra semicolon can be omitted. GDB (at least through version 4.9) doesn't support any kind of index type other than a range anyway; I'm not sure about dbx.
It is well established, and widely used, that the type of the index, unlike most types found in the stabs, is merely a type definition, not type information (see section The String Field) (that is, it need not start with `type-number=' if it is defining a new type). According to a comment in GDB, this is also true of the type of the array elements; it gives `ar1;1;10;ar1;1;10;4' as a legitimate way to express a two dimensional array. According to AIX documentation, the element type must be type information. GDB accepts either.
The type of the index is often a range type, expressed as the type descriptor `r' and some parameters. It defines the size of the array. In the example below, the range `r1;0;2;' defines an index type which is a subrange of type 1 (integer), with a lower bound of 0 and an upper bound of 2. This defines the valid range of subscripts of a three-element C array.
For example, the definition:
char char_vec[3] = {'a','b','c'};
produces the output:
.stabs "char_vec:G19=ar1;0;2;2",32,0,0,0
.global _char_vec
.align 4
_char_vec:
.byte 97
.byte 98
.byte 99
If an array is packed, the elements are spaced more closely than normal, saving memory at the expense of speed. For example, an array of 3-byte objects might, if unpacked, have each element aligned on a 4-byte boundary, but if packed, have no padding. One way to specify that something is packed is with type attributes (see section The String Field). In the case of arrays, another is to use the `P' type descriptor instead of `a'. Other than specifying a packed array, `P' is identical to `a'.
An open array is represented by the `A' type descriptor followed by type information specifying the type of the array elements.
An N-dimensional dynamic array is represented by
D dimensions ; type-information
dimensions is the number of dimensions; type-information specifies the type of the array elements.
A subarray of an N-dimensional array is represented by
E dimensions ; type-information
dimensions is the number of dimensions; type-information specifies the type of the array elements.
Some languages, like C or the original Pascal, do not have string types, they just have related things like arrays of characters. But most Pascals and various other languages have string types, which are indicated as follows:
n type-information ; bytes
z type-information ; bytes
N
Languages, such as CHILL which have a string type which is basically just an array of characters use the `S' type attribute (see section The String Field).
Enumerations are defined with the `e' type descriptor.
The source line below declares an enumeration type at file scope.
The type definition is located after the N_RBRAC that marks the end of
the previous procedure's block scope, and before the N_FUN that marks
the beginning of the next procedure's block scope. Therefore it does not
describe a block local symbol, but a file local one.
The source line:
enum e_places {first,second=3,last};
generates the following stab:
.stabs "e_places:T22=efirst:0,second:3,last:4,;",128,0,0,0
The symbol descriptor (`T') says that the stab describes a structure, enumeration, or union tag. The type descriptor `e', following the `22=' of the type definition narrows it down to an enumeration type. Following the `e' is a list of the elements of the enumeration. The format is `name:value,'. The list of elements ends with `;'. The fact that value is specified as an integer can cause problems if the value is large. GCC 2.5.2 tries to output it in octal in that case with a leading zero, which is probably a good thing, although GDB 4.11 supports octal only in cases where decimal is perfectly good. Negative decimal values are supported by both GDB and dbx.
There is no standard way to specify the size of an enumeration type; it is determined by the architecture (normally all enumerations types are 32 bits). Type attributes can be used to specify an enumeration type of another size for debuggers which support them; see section The String Field.
Enumeration types are unusual in that they define symbols for the
enumeration values (first, second, and third in the
above example), and even though these symbols are visible in the file as
a whole (rather than being in a more local namespace like structure
member names), they are defined in the type definition for the
enumeration type rather than each having their own symbol. In order to
be fast, GDB will only get symbols from such types (in its initial scan
of the stabs) if the type is the first thing defined after a `T' or
`t' symbol descriptor (the above example fulfills this
requirement). If the type does not have a name, the compiler should
emit it in a nameless stab (see section The String Field); GCC does this.
The encoding of structures in stabs can be shown with an example.
The following source code declares a structure tag and defines an
instance of the structure in global scope. Then a typedef equates the
structure tag with a new type. Seperate stabs are generated for the
structure tag, the structure typedef, and the structure instance. The
stabs for the tag and the typedef are emited when the definitions are
encountered. Since the structure elements are not initialized, the
stab and code for the structure variable itself is located at the end
of the program in the bss section.
struct s_tag {
int s_int;
float s_float;
char s_char_vec[8];
struct s_tag* s_next;
} g_an_s;
typedef struct s_tag s_typedef;
The structure tag has an N_LSYM stab type because, like the
enumeration, the symbol has file scope. Like the enumeration, the
symbol descriptor is `T', for enumeration, structure, or tag type.
The type descriptor `s' following the `16=' of the type
definition narrows the symbol type to structure.
Following the `s' type descriptor is the number of bytes the structure occupies, followed by a description of each structure element. The structure element descriptions are of the form name:type, bit offset from the start of the struct, number of bits in the element.
# 128 is N_LSYM
.stabs "s_tag:T16=s20s_int:1,0,32;s_float:12,32,32;
s_char_vec:17=ar1;0;7;2,64,64;s_next:18=*16,128,32;;",128,0,0,0
In this example, the first two structure elements are previously defined
types. For these, the type following the `name:' part of the
element description is a simple type reference. The other two structure
elements are new types. In this case there is a type definition
embedded after the `name:'. The type definition for the
array element looks just like a type definition for a standalone array.
The s_next field is a pointer to the same kind of structure that
the field is an element of. So the definition of structure type 16
contains a type definition for an element which is a pointer to type 16.
If a field is a static member (this is a C++ feature in which a single variable appears to be a field of every structure of a given type) it still starts out with the field name, a colon, and the type, but then instead of a comma, bit position, comma, and bit size, there is a colon followed by the name of the variable which each such field refers to.
If the structure has methods (a C++ feature), they follow the non-method fields; see section GNU C++ Stabs.
To give a type a name, use the `t' symbol descriptor. The type is specified by the type information (see section The String Field) for the stab. For example,
.stabs "s_typedef:t16",128,0,0,0 # 128 is N_LSYM
specifies that s_typedef refers to type number 16. Such stabs
have symbol type N_LSYM (or C_DECL for XCOFF).
If you are specifying the tag name for a structure, union, or enumeration, use the `T' symbol descriptor instead. I believe C is the only language with this feature.
If the type is an opaque type (I believe this is a Modula-2 feature), AIX provides a type descriptor to specify it. The type descriptor is `o' and is followed by a name. I don't know what the name means--is it always the same as the name of the type, or is this type descriptor used with a nameless stab (see section The String Field)? There optionally follows a comma followed by type information which defines the type of this type. If omitted, a semicolon is used in place of the comma and the type information, and the type is much like a generic pointer type--it has a known size but little else about it is specified.
union u_tag {
int u_int;
float u_float;
char* u_char;
} an_u;
This code generates a stab for a union tag and a stab for a union
variable. Both use the N_LSYM stab type. If a union variable is
scoped locally to the procedure in which it is defined, its stab is
located immediately preceding the N_LBRAC for the procedure's block
start.
The stab for the union tag, however, is located preceding the code for
the procedure in which it is defined. The stab type is N_LSYM. This
would seem to imply that the union type is file scope, like the struct
type s_tag. This is not true. The contents and position of the stab
for u_type do not convey any infomation about its procedure local
scope.
# 128 is N_LSYM
.stabs "u_tag:T23=u4u_int:1,0,32;u_float:12,0,32;u_char:21,0,32;;",
128,0,0,0
The symbol descriptor `T', following the `name:' means that the stab describes an enumeration, structure, or union tag. The type descriptor `u', following the `23=' of the type definition, narrows it down to a union type definition. Following the `u' is the number of bytes in the union. After that is a list of union element descriptions. Their format is name:type, bit offset into the union, number of bytes for the element;.
The stab for the union variable is:
.stabs "an_u:23",128,0,0,-20 # 128 is N_LSYM
`-20' specifies where the variable is stored (see section Automatic Variables Allocated on the Stack).
Various types can be defined for function variables. These types are not used in defining functions (see section Procedures); they are used for things like pointers to functions.
The simple, traditional, type is type descriptor `f' is followed by type information for the return type of the function, followed by a semicolon.
This does not deal with functions for which the number and types of the parameters are part of the type, as in Modula-2 or ANSI C. AIX provides extensions to specify these, using the `f', `F', `p', and `R' type descriptors.
First comes the type descriptor. If it is `f' or `F', this type involves a function rather than a procedure, and the type information for the return type of the function follows, followed by a comma. Then comes the number of parameters to the function and a semicolon. Then, for each parameter, there is the name of the parameter followed by a colon (this is only present for type descriptors `R' and `F' which represent Pascal function or procedure parameters), type information for the parameter, a comma, 0 if passed by reference or 1 if passed by value, and a semicolon. The type definition ends with a semicolon.
For example, this variable definition:
int (*g_pf)();
generates the following code:
.stabs "g_pf:G24=*25=f1",32,0,0,0
.common _g_pf,4,"bss"
The variable defines a new type, 24, which is a pointer to another new
type, 25, which is a function returning int.
This chapter describes the format of symbol table entries and how stab assembler directives map to them. It also describes the transformations that the assembler and linker make on data from stabs.
Each time the assembler encounters a stab directive, it puts each field of the stab into a corresponding field in a symbol table entry of its output file. If the stab contains a string field, the symbol table entry for that stab points to a string table entry containing the string data from the stab. Assembler labels become relocatable addresses. Symbol table entries in a.out have the format:
struct internal_nlist {
unsigned long n_strx; /* index into string table of name */
unsigned char n_type; /* type of symbol */
unsigned char n_other; /* misc info (usually empty) */
unsigned short n_desc; /* description field */
bfd_vma n_value; /* value of symbol */
};
If the stab has a string, the n_strx field holds the offset in
bytes of the string within the string table. The string is terminated
by a NUL character. If the stab lacks a string (for example, it was
produced by a .stabn or .stabd directive), the
n_strx field is zero.
Symbol table entries with n_type field values greater than 0x1f
originated as stabs generated by the compiler (with one random
exception). The other entries were placed in the symbol table of the
executable by the assembler or the linker.
The linker concatenates object files and does fixups of externally defined symbols.
You can see the transformations made on stab data by the assembler and linker by examining the symbol table after each pass of the build. To do this, use `nm -ap', which dumps the symbol table, including debugging information, unsorted. For stab entries the columns are: value, other, desc, type, string. For assembler and linker symbols, the columns are: value, type, string.
The low 5 bits of the stab type tell the linker how to relocate the
value of the stab. Thus for stab types like N_RSYM and
N_LSYM, where the value is an offset or a register number, the
low 5 bits are N_ABS, which tells the linker not to relocate the
value.
Where the value of a stab contains an assembly language label, it is transformed by each build step. The assembler turns it into a relocatable address and the linker turns it into an absolute address.
This source line defines a static variable at file scope:
static int s_g_repeat
The following stab describes the symbol:
.stabs "s_g_repeat:S1",38,0,0,_s_g_repeat
The assembler transforms the stab into this symbol table entry in the `.o' file. The location is expressed as a data segment offset.
00000084 - 00 0000 STSYM s_g_repeat:S1
In the symbol table entry from the executable, the linker has made the relocatable address absolute.
0000e00c - 00 0000 STSYM s_g_repeat:S1
Stabs for global variables do not contain location information. In this case, the debugger finds location information in the assembler or linker symbol table entry describing the variable. The source line:
char g_foo = 'c';
generates the stab:
.stabs "g_foo:G2",32,0,0,0
The variable is represented by two symbol table entries in the object
file (see below). The first one originated as a stab. The second one
is an external symbol. The upper case `D' signifies that the
n_type field of the symbol table contains 7, N_DATA with
local linkage. The stab's value is zero since the value is not used for
N_GSYM stabs. The value of the linker symbol is the relocatable
address corresponding to the variable.
00000000 - 00 0000 GSYM g_foo:G2 00000080 D _g_foo
These entries as transformed by the linker. The linker symbol table entry now holds an absolute address:
00000000 - 00 0000 GSYM g_foo:G2 ... 0000e008 D _g_foo
For object file formats using stabs in separate sections (see section Using Stabs in Their Own Sections), use objdump --stabs instead of nm to show the
stabs in an object or executable file. objdump is a GNU utility;
Sun does not provide any equivalent.
The following example is for a stab whose value is an address is relative to the compilation unit (see section Having the Linker Relocate Stabs in ELF). For example, if the source line
static int ld = 5;
appears within a function, then the assembly language output from the compiler contains:
.Ddata.data:
...
.stabs "ld:V(0,3)",0x26,0,4,.L18-Ddata.data # 0x26 is N_STSYM
...
.L18:
.align 4
.word 0x5
Because the value is formed by subtracting one symbol from another, the value is absolute, not relocatable, and so the object file contains
Symnum n_type n_othr n_desc n_value n_strx String 31 STSYM 0 4 00000004 680 ld:V(0,3)
without any relocations, and the executable file also contains
Symnum n_type n_othr n_desc n_value n_strx String 31 STSYM 0 4 00000004 680 ld:V(0,3)
In C++, a class name which is declared with class, struct,
or union, is not only a tag, as in C, but also a type name. Thus
there should be stabs with both `t' and `T' symbol descriptors
(see section Giving a Type a Name).
To save space, there is a special abbreviation for this case. If the `T' symbol descriptor is followed by `t', then the stab defines both a type name and a tag.
For example, the C++ code
struct foo {int x;};
can be represented as either
.stabs "foo:T19=s4x:1,0,32;;",128,0,0,0 # 128 is N_LSYM .stabs "foo:t19",128,0,0,0
or
.stabs "foo:Tt19=s4x:1,0,32;;",128,0,0,0
In C++, a symbol (such as a type name) can be defined within another type.
In stabs, this is sometimes represented by making the name of a symbol which contains `::'. Such a pair of colons does not end the name of the symbol, the way a single colon would (see section The String Field). I'm not sure how consistently used or well thought out this mechanism is. So that a pair of colons in this position always has this meaning, `:' cannot be used as a symbol descriptor.
For example, if the string for a stab is `foo::bar::baz:t5=*6',
then foo::bar::baz is the name of the symbol, `t' is the
symbol descriptor, and `5=*6' is the type information.
<< the examples that follow are based on a01.C >>
C++ adds two more builtin types to the set defined for C. These are the unknown type and the vtable record type. The unknown type, type 16, is defined in terms of itself like the void type.
The vtable record type, type 17, is defined as a structure type and then as a structure tag. The structure has four fields: delta, index, pfn, and delta2. pfn is the function pointer.
<< In boilerplate $vtbl_ptr_type, what are the fields delta, index, and delta2 used for? >>
This basic type is present in all C++ programs even if there are no virtual methods defined.
.stabs "struct_name:sym_desc(type)type_def(17)=type_desc(struct)struct_bytes(8)
elem_name(delta):type_ref(short int),bit_offset(0),field_bits(16);
elem_name(index):type_ref(short int),bit_offset(16),field_bits(16);
elem_name(pfn):type_def(18)=type_desc(ptr to)type_ref(void),
bit_offset(32),field_bits(32);
elem_name(delta2):type_def(short int);bit_offset(32),field_bits(16);;"
N_LSYM, NIL, NIL
.stabs "$vtbl_ptr_type:t17=s8
delta:6,0,16;index:6,16,16;pfn:18=*15,32,32;delta2:6,32,16;;"
,128,0,0,0
.stabs "name:sym_dec(struct tag)type_ref($vtbl_ptr_type)",N_LSYM,NIL,NIL,NIL
.stabs "$vtbl_ptr_type:T17",128,0,0,0
The stabs describing C++ language features are an extension of the stabs describing C. Stabs representing C++ class types elaborate extensively on the stab format used to describe structure types in C. Stabs representing class type variables look just like stabs representing C language variables.
Consider the following very simple class definition.
class baseA {
public:
int Adat;
int Ameth(int in, char other);
};
The class baseA is represented by two stabs. The first stab describes
the class as a structure type. The second stab describes a structure
tag of the class type. Both stabs are of stab type N_LSYM. Since the
stab is not located between an N_FUN and an N_LBRAC stab this indicates
that the class is defined at file scope. If it were, then the N_LSYM
would signify a local variable.
A stab describing a C++ class type is similar in format to a stab describing a C struct, with each class member shown as a field in the structure. The part of the struct format describing fields is expanded to include extra information relevent to C++ class members. In addition, if the class has multiple base classes or virtual functions the struct format outside of the field parts is also augmented.
In this simple example the field part of the C++ class stab representing member data looks just like the field part of a C struct stab. The section on protections describes how its format is sometimes extended for member data.
The field part of a C++ class stab representing a member function differs substantially from the field part of a C struct stab. It still begins with `name:' but then goes on to define a new type number for the member function, describe its return type, its argument types, its protection level, any qualifiers applied to the method definition, and whether the method is virtual or not. If the method is virtual then the method description goes on to give the vtable index of the method, and the type number of the first base class defining the method.
When the field name is a method name it is followed by two colons rather than one. This is followed by a new type definition for the method. This is a number followed by an equal sign and the type of the method. Normally this will be a type declared using the `#' type descriptor; see section The `#' Type Descriptor; static member functions are declared using the `f' type descriptor instead; see section Function Types.
The format of an overloaded operator method name differs from that of other methods. It is `op$::operator-name.' where operator-name is the operator name such as `+' or `+='. The name ends with a period, and any characters except the period can occur in the operator-name string.
The next part of the method description represents the arguments to the
method, preceeded by a colon and ending with a semi-colon. The types of
the arguments are expressed in the same way argument types are expressed
in C++ name mangling. In this example an int and a char
map to `ic'.
This is followed by a number, a letter, and an asterisk or period, followed by another semicolon. The number indicates the protections that apply to the member function. Here the 2 means public. The letter encodes any qualifier applied to the method definition. In this case, `A' means that it is a normal function definition. The dot shows that the method is not virtual. The sections that follow elaborate further on these fields and describe the additional information present for virtual methods.
.stabs "class_name:sym_desc(type)type_def(20)=type_desc(struct)struct_bytes(4)
field_name(Adat):type(int),bit_offset(0),field_bits(32);
method_name(Ameth)::type_def(21)=type_desc(method)return_type(int);
:arg_types(int char);
protection(public)qualifier(normal)virtual(no);;"
N_LSYM,NIL,NIL,NIL
.stabs "baseA:t20=s4Adat:1,0,32;Ameth::21=##1;:ic;2A.;;",128,0,0,0 .stabs "class_name:sym_desc(struct tag)",N_LSYM,NIL,NIL,NIL .stabs "baseA:T20",128,0,0,0
As shown above, describing even a simple C++ class definition is accomplished by massively extending the stab format used in C to describe structure types. However, once the class is defined, C stabs with no modifications can be used to describe class instances. The following source:
main () {
baseA AbaseA;
}
yields the following stab describing the class instance. It looks no different from a standard C stab describing a local variable.
.stabs "name:type_ref(baseA)", N_LSYM, NIL, NIL, frame_ptr_offset
.stabs "AbaseA:20",128,0,0,-20
The class definition shown above declares Ameth. The C++ source below defines Ameth:
int
baseA::Ameth(int in, char other)
{
return in;
};
This method definition yields three stabs following the code of the
method. One stab describes the method itself and following two describe
its parameters. Although there is only one formal argument all methods
have an implicit argument which is the this pointer. The this
pointer is a pointer to the object on which the method was called. Note
that the method name is mangled to encode the class name and argument
types. Name mangling is described in the ARM (The Annotated
C++ Reference Manual, by Ellis and Stroustrup, ISBN
0-201-51459-1); `gpcompare.texi' in Cygnus GCC distributions
describes the differences between GNU mangling and ARM
mangling.
.stabs "name:symbol_desriptor(global function)return_type(int)",
N_FUN, NIL, NIL, code_addr_of_method_start
.stabs "Ameth__5baseAic:F1",36,0,0,_Ameth__5baseAic
Here is the stab for the this pointer implicit argument. The
name of the this pointer is always this. Type 19, the
this pointer is defined as a pointer to type 20, baseA,
but a stab defining baseA has not yet been emited. Since the
compiler knows it will be emited shortly, here it just outputs a cross
reference to the undefined symbol, by prefixing the symbol name with
`xs'.
.stabs "name:sym_desc(register param)type_def(19)=
type_desc(ptr to)type_ref(baseA)=
type_desc(cross-reference to)baseA:",N_RSYM,NIL,NIL,register_number
.stabs "this:P19=*20=xsbaseA:",64,0,0,8
The stab for the explicit integer argument looks just like a parameter to a C function. The last field of the stab is the offset from the argument pointer, which in most systems is the same as the frame pointer.
.stabs "name:sym_desc(value parameter)type_ref(int)",
N_PSYM,NIL,NIL,offset_from_arg_ptr
.stabs "in:p1",160,0,0,72
<< The examples that follow are based on A1.C >>
This is like the `f' type descriptor for functions (see section Function Types), except that a function which uses the `#' type descriptor
takes an extra argument as its first argument, for the this
pointer. The `#' type descriptor is optionally followed by the
types of the arguments, then another `#'. If the types of the
arguments are omitted, so that the second `#' immediately follows
the `#' which is the type descriptor, the arguments are being
omitted (to save space) and can be deduced from the mangled name of the
method. After the second `#' there is type information for the
return type of the method and a semicolon.
Note that although such a type will normally be used to describe fields in structures, unions, or classes, for at least some versions of the compiler it can also be used in other contexts.
The `@' type descriptor is for a member (class and variable) type. It is followed by type information for the offset basetype, a comma, and type information for the type of the field being pointed to. (FIXME: this is acknowledged to be gibberish. Can anyone say what really goes here?).
Note that there is a conflict between this and type attributes (see section The String Field); both use type descriptor `@'. Fortunately, the `@' type descriptor used in this C++ sense always will be followed by a digit, `(', or `-', and type attributes never start with those things.
In the simple class definition shown above all member data and functions were publicly accessable. The example that follows contrasts public, protected and privately accessable fields and shows how these protections are encoded in C++ stabs.
If the character following the `field-name:' part of the string is `/', then the next character is the visibility. `0' means private, `1' means protected, and `2' means public. Debuggers should ignore visibility characters they do not recognize, and assume a reasonable default (such as public) (GDB 4.11 does not, but this should be fixed in the next GDB release). If no visibility is specified the field is public. The visibility `9' means that the field has been optimized out and is public (there is no way to specify an optimized out field with a private or protected visibility). Visibility `9' is not supported by GDB 4.11; this should be fixed in the next GDB release.
The following C++ source:
class vis {
private:
int priv;
protected:
char prot;
public:
float pub;
};
generates the following stab:
# 128 is N_LSYM .stabs "vis:T19=s12priv:/01,0,32;prot:/12,32,8;pub:12,64,32;;",128,0,0,0
`vis:T19=s12' indicates that type number 19 is a 12 byte structure
named vis The priv field has public visibility
(`/0'), type int (`1'), and offset and size `,0,32;'.
The prot field has protected visibility (`/1'), type char
(`2') and offset and size `,32,8;'. The pub field has
type float (`12'), and offset and size `,64,32;'.
Protections for member functions are signified by one digit embeded in the field part of the stab describing the method. The digit is 0 if private, 1 if protected and 2 if public. Consider the C++ class definition below:
class all_methods {
private:
int priv_meth(int in){return in;};
protected:
char protMeth(char in){return in;};
public:
float pubMeth(float in){return in;};
};
It generates the following stab. The digit in question is to the left of an `A' in each case. Notice also that in this case two symbol descriptors apply to the class name struct tag and struct type.
.stabs "class_name:sym_desc(struct tag&type)type_def(21)=
sym_desc(struct)struct_bytes(1)
meth_name::type_def(22)=sym_desc(method)returning(int);
:args(int);protection(private)modifier(normal)virtual(no);
meth_name::type_def(23)=sym_desc(method)returning(char);
:args(char);protection(protected)modifier(normal)virual(no);
meth_name::type_def(24)=sym_desc(method)returning(float);
:args(float);protection(public)modifier(normal)virtual(no);;",
N_LSYM,NIL,NIL,NIL
.stabs "all_methods:Tt21=s1priv_meth::22=##1;:i;0A.;protMeth::23=##2;:c;1A.;
pubMeth::24=##12;:f;2A.;;",128,0,0,0
const, volatile, const volatile)<< based on a6.C >>
In the class example described above all the methods have the normal modifier. This method modifier information is located just after the protection information for the method. This field has four possible character values. Normal methods use `A', const methods use `B', volatile methods use `C', and const volatile methods use `D'. Consider the class definition below:
class A {
public:
int ConstMeth (int arg) const { return arg; };
char VolatileMeth (char arg) volatile { return arg; };
float ConstVolMeth (float arg) const volatile {return arg; };
};
This class is described by the following stab:
.stabs "class(A):sym_desc(struct)type_def(20)=type_desc(struct)struct_bytes(1)
meth_name(ConstMeth)::type_def(21)sym_desc(method)
returning(int);:arg(int);protection(public)modifier(const)virtual(no);
meth_name(VolatileMeth)::type_def(22)=sym_desc(method)
returning(char);:arg(char);protection(public)modifier(volatile)virt(no)
meth_name(ConstVolMeth)::type_def(23)=sym_desc(method)
returning(float);:arg(float);protection(public)modifer(const volatile)
virtual(no);;", ...
.stabs "A:T20=s1ConstMeth::21=##1;:i;2B.;VolatileMeth::22=##2;:c;2C.;
ConstVolMeth::23=##12;:f;2D.;;",128,0,0,0
<< The following examples are based on a4.C >>
The presence of virtual methods in a class definition adds additional data to the class description. The extra data is appended to the description of the virtual method and to the end of the class description. Consider the class definition below:
class A {
public:
int Adat;
virtual int A_virt (int arg) { return arg; };
};
This results in the stab below describing class A. It defines a new type (20) which is an 8 byte structure. The first field of the class struct is `Adat', an integer, starting at structure offset 0 and occupying 32 bits.
The second field in the class struct is not explicitly defined by the C++ class definition but is implied by the fact that the class contains a virtual method. This field is the vtable pointer. The name of the vtable pointer field starts with `$vf' and continues with a type reference to the class it is part of. In this example the type reference for class A is 20 so the name of its vtable pointer field is `$vf20', followed by the usual colon.
Next there is a type definition for the vtable pointer type (21). This is in turn defined as a pointer to another new type (22).
Type 22 is the vtable itself, which is defined as an array, indexed by a range of integers between 0 and 1, and whose elements are of type 17. Type 17 was the vtable record type defined by the boilerplate C++ type definitions, as shown earlier.
The bit offset of the vtable pointer field is 32. The number of bits in the field are not specified when the field is a vtable pointer.
Next is the method definition for the virtual member function A_virt.
Its description starts out using the same format as the non-virtual
member functions described above, except instead of a dot after the
`A' there is an asterisk, indicating that the function is virtual.
Since is is virtual some addition information is appended to the end
of the method description.
The first number represents the vtable index of the method. This is a 32 bit unsigned number with the high bit set, followed by a semi-colon.
The second number is a type reference to the first base class in the inheritence hierarchy defining the virtual member function. In this case the class stab describes a base class so the virtual function is not overriding any other definition of the method. Therefore the reference is to the type number of the class that the stab is describing (20).
This is followed by three semi-colons. One marks the end of the current sub-section, one marks the end of the method field, and the third marks the end of the struct definition.
For classes containing virtual functions the very last section of the string part of the stab holds a type reference to the first base class. This is preceeded by `~%' and followed by a final semi-colon.
.stabs "class_name(A):type_def(20)=sym_desc(struct)struct_bytes(8)
field_name(Adat):type_ref(int),bit_offset(0),field_bits(32);
field_name(A virt func ptr):type_def(21)=type_desc(ptr to)type_def(22)=
sym_desc(array)index_type_ref(range of int from 0 to 1);
elem_type_ref(vtbl elem type),
bit_offset(32);
meth_name(A_virt)::typedef(23)=sym_desc(method)returning(int);
:arg_type(int),protection(public)normal(yes)virtual(yes)
vtable_index(1);class_first_defining(A);;;~%first_base(A);",
N_LSYM,NIL,NIL,NIL
.stabs "A:t20=s8Adat:1,0,32;$vf20:21=*22=ar1;0;1;17,32;
A_virt::23=##1;:i;2A*-2147483647;20;;;~%20;",128,0,0,0
Stabs describing C++ derived classes include additional sections that describe the inheritence hierarchy of the class. A derived class stab also encodes the number of base classes. For each base class it tells if the base class is virtual or not, and if the inheritence is private or public. It also gives the offset into the object of the portion of the object corresponding to each base class.
This additional information is embeded in the class stab following the number of bytes in the struct. First the number of base classes appears bracketed by an exclamation point and a comma.
Then for each base type there repeats a series: a virtual character, a visibilty character, a number, a comma, another number, and a semi-colon.
The virtual character is `1' if the base class is virtual and `0' if not. The visibility character is `2' if the derivation is public, `1' if it is protected, and `0' if it is private. Debuggers should ignore virtual or visibility characters they do not recognize, and assume a reasonable default (such as public and non-virtual) (GDB 4.11 does not, but this should be fixed in the next GDB release).
The number following the virtual and visibility characters is the offset from the start of the object to the part of the object pertaining to the base class.
After the comma, the second number is a type_descriptor for the base type. Finally a semi-colon ends the series, which repeats for each base class.
The source below defines three base classes A, B, and
C and the derived class D.
class A {
public:
int Adat;
virtual int A_virt (int arg) { return arg; };
};
class B {
public:
int B_dat;
virtual int B_virt (int arg) {return arg; };
};
class C {
public:
int Cdat;
virtual int C_virt (int arg) {return arg; };
};
class D : A, virtual B, public C {
public:
int Ddat;
virtual int A_virt (int arg ) { return arg+1; };
virtual int B_virt (int arg) { return arg+2; };
virtual int C_virt (int arg) { return arg+3; };
virtual int D_virt (int arg) { return arg; };
};
Class stabs similar to the ones described earlier are generated for each base class.
.stabs "A:T20=s8Adat:1,0,32;$vf20:21=*22=ar1;0;1;17,32;
A_virt::23=##1;:i;2A*-2147483647;20;;;~%20;",128,0,0,0
.stabs "B:Tt25=s8Bdat:1,0,32;$vf25:21,32;B_virt::26=##1;
:i;2A*-2147483647;25;;;~%25;",128,0,0,0
.stabs "C:Tt28=s8Cdat:1,0,32;$vf28:21,32;C_virt::29=##1;
:i;2A*-2147483647;28;;;~%28;",128,0,0,0
In the stab describing derived class D below, the information about
the derivation of this class is encoded as follows.
.stabs "derived_class_name:symbol_descriptors(struct tag&type)=
type_descriptor(struct)struct_bytes(32)!num_bases(3),
base_virtual(no)inheritence_public(no)base_offset(0),
base_class_type_ref(A);
base_virtual(yes)inheritence_public(no)base_offset(NIL),
base_class_type_ref(B);
base_virtual(no)inheritence_public(yes)base_offset(64),
base_class_type_ref(C); ...
.stabs "D:Tt31=s32!3,000,20;100,25;0264,28;$vb25:24,128;Ddat:
1,160,32;A_virt::32=##1;:i;2A*-2147483647;20;;B_virt:
:32:i;2A*-2147483647;25;;C_virt::32:i;2A*-2147483647;
28;;D_virt::32:i;2A*-2147483646;31;;;~%20;",128,0,0,0
A derived class object consists of a concatination in memory of the data
areas defined by each base class, starting with the leftmost and ending
with the rightmost in the list of base classes. The exception to this
rule is for virtual inheritence. In the example above, class D
inherits virtually from base class B. This means that an
instance of a D object will not contain its own B part but
merely a pointer to a B part, known as a virtual base pointer.
In a derived class stab, the base offset part of the derivation
information, described above, shows how the base class parts are
ordered. The base offset for a virtual base class is always given as 0.
Notice that the base offset for B is given as 0 even though
B is not the first base class. The first base class A
starts at offset 0.
The field information part of the stab for class D describes the field
which is the pointer to the virtual base class B. The vbase pointer
name is `$vb' followed by a type reference to the virtual base class.
Since the type id for B in this example is 25, the vbase pointer name
is `$vb25'.
.stabs "D:Tt31=s32!3,000,20;100,25;0264,28;$vb25:24,128;Ddat:1,
160,32;A_virt::32=##1;:i;2A*-2147483647;20;;B_virt::32:i;
2A*-2147483647;25;;C_virt::32:i;2A*-2147483647;28;;D_virt:
:32:i;2A*-2147483646;31;;;~%20;",128,0,0,0
Following the name and a semicolon is a type reference describing the
type of the virtual base class pointer, in this case 24. Type 24 was
defined earlier as the type of the B class this pointer. The
this pointer for a class is a pointer to the class type.
.stabs "this:P24=*25=xsB:",64,0,0,8
Finally the field offset part of the vbase pointer field description
shows that the vbase pointer is the first field in the D object,
before any data fields defined by the class. The layout of a D
class object is a follows, Adat at 0, the vtable pointer for
A at 32, Cdat at 64, the vtable pointer for C at 96, the
virtual base pointer for B at 128, and Ddat at 160.
The data area for a class is a concatenation of the space used by the data members of the class. If the class has virtual methods, a vtable pointer follows the class data. The field offset part of each field description in the class stab shows this ordering.
<< How is this reflected in stabs? See Cygnus bug #677 for some info. >>
The following are all the possible values for the stab type field, for a.out files, in numeric order. This does not apply to XCOFF, but it does apply to stabs in sections (see section Using Stabs in Their Own Sections). Stabs in ECOFF use these values but add 0x8f300 to distinguish them from non-stab symbols.
The symbolic names are defined in the file `include/aout/stabs.def'.
The following types are used by the linker and assembler, not by stab directives. Since this document does not attempt to describe aspects of object file format other than the debugging format, no details are given.
0x0 N_UNDF
0x2 N_ABS
0x3 N_ABS | N_EXT
0x4 N_TEXT
0x5 N_TEXT | N_EXT
0x6 N_DATA
0x7 N_DATA | N_EXT
0x8 N_BSS
0x9 N_BSS | N_EXT
0x0c N_FN_SEQ
N_FN, for Sequent compilers
0x0a N_INDR
0x12 N_COMM
0x14 N_SETA
0x15 N_SETA | N_EXT
0x16 N_SETT
0x17 N_SETT | N_EXT
0x18 N_SETD
0x19 N_SETD | N_EXT
0x1a N_SETB
0x1b N_SETB | N_EXT
0x1c N_SETV
0x1d N_SETV | N_EXT
0x1e N_WARNING
0x1f N_FN
The following symbol types indicate that this is a stab. This is the full list of stab numbers, including stab types that are used in languages other than C.
0x20 N_GSYM
0x22 N_FNAME
0x24 N_FUN
0x26 N_STSYM
0x28 N_LCSYM
0x2a N_MAIN
0x2c N_ROSYM
.rodata section; see section Static Variables.
0x30 N_PC
0x32 N_NSYMS
0x34 N_NOMAP
0x38 N_OBJ
0x3c N_OPT
0x40 N_RSYM
0x42 N_M2C
0x44 N_SLINE
0x46 N_DSLINE
0x48 N_BSLINE
0x48 N_BROWS
0x4a N_DEFD
0x4c N_FLINE
0x50 N_EHDECL
0x50 N_MOD2
0x54 N_CATCH
catch clause; see section N_CATCH.
0x60 N_SSYM
0x62 N_ENDM
0x64 N_SO
0x80 N_LSYM
0x82 N_BINCL
0x84 N_SOL
0xa0 N_PSYM
0xa2 N_EINCL
0xa4 N_ENTRY
0xc0 N_LBRAC
0xc2 N_EXCL
0xc4 N_SCOPE
0xe0 N_RBRAC
0xe2 N_BCOMM
0xe4 N_ECOMM
0xe8 N_ECOML
0xea N_WITH
with statement: type,,0,0,offset (Solaris2).
0xf0 N_NBTEXT
0xf2 N_NBDATA
0xf4 N_NBBSS
0xf6 N_NBSTS
0xf8 N_NBLCS
The symbol descriptor is the character which follows the colon in many stabs, and which tells what kind of stab it is. See section The String Field, for more information about their use.
digit
(
-
:
a
b
c
C
N_CATCH and the former uses
another symbol type.
d
D
f
F
G
i
I
J
L
m
p
pP
pF
P
N_PSYM); see section Parameters. Prototype of function
referenced by this file (Sun acc) (symbol type N_FUN).
Q
R
r
S
s
t
T
v
V
x
X
The type descriptor is the character which follows the type number and an equals sign. It specifies what kind of type is being defined. See section The String Field, for more information about their use.
digit
(
-
#
*
&
@
a
A
b
B
c
C
d
D
e
E
f
F
g
G
i
k
K
M
n
N
o
p
P
r
R
s
S
u
v
w
x
Y
z
For a full list of stab types, and cross-references to where they are described, see section Table of Stab Types. This appendix just covers certain stabs which are not yet described in the main body of this document; eventually the information will all be in one place.
Format of an entry:
The first line is the symbol type (see `include/aout/stab.def').
The second line describes the language constructs the symbol type represents.
The third line is the stab format with the significant stab fields named and the rest NIL.
Subsequent lines expand upon the meaning and possible values for each significant stab field.
Finally, any further information.
.stabs: N_PC
"name" -> "symbol_name" <<?>> value -> supposedly the line number (stab.def is skeptical)
`stabdump.c' says: global pascal symbol: name,,0,subtype,line << subtype? >>
.stabn: N_NSYMS
0, files,,funcs,lines (stab.def)
.stabs: N_NOMAP
name, ,0,type,ignored (stab.def)
.stabs: N_M2C
"string" -> "unit_name,unit_time_stamp[,code_time_stamp]"
desc -> unit_number
value -> 0 (main unit)
1 (any other unit)
See Dbx and Dbxtool Interfaces, 2nd edition, by Sun, 1988, for more information.
.stabs: N_BROWS
<<?>> "path to associated `.cb' file"
Note: N_BROWS has the same value as N_BSLINE.
.stabn: N_DEFD
GNU Modula-2 definition module dependency. The value is the
modification time of the definition file. The other field is non-zero
if it is imported with the GNU M2 keyword %INITIALIZE. Perhaps
N_M2C can be used if there are enough empty fields?
.stabs: N_EHDECL
"string is variable name"
Note: conflicts with N_MOD2.
.stab?: N_MOD2
Note: conflicts with N_EHDECL <<?>>
.stabn: N_CATCH
catch clause
GNU C++ catch clause. The value is its address. The desc field
is nonzero if this entry is immediately followed by a CAUGHT stab
saying what exception was caught. Multiple CAUGHT stabs means
that multiple exceptions can be caught here. If desc is 0, it means all
exceptions are caught here.
.stabn: N_SSYM
The value is the offset in the structure.
<<?looking at structs and unions in C I didn't see these>>
.stab?: N_SCOPE
.stab?: N_NBTEXT
.stab?: N_NBDATA
.stab?: N_NBBSS
.stab?: N_NBSTS
.stab?: N_NBLCS
However, the following values are not the values used by Gould; they are the values which GNU has been documenting for these values for a long time, without actually checking what Gould uses. I include these values only because perhaps some someone actually did something with the GNU information (I hope not, why GNU knowingly assigned wrong values to these in the header file is a complete mystery to me).
240 0xf0 N_NBTEXT ?? 242 0xf2 N_NBDATA ?? 244 0xf4 N_NBBSS ?? 246 0xf6 N_NBSTS ?? 248 0xf8 N_NBLCS ??
.stabn: N_LENG
N_LSYM and
N_GSYM), the desc field is supposed to contain the source
line number on which the variable is defined. In reality the desc
field is always 0. (This behavior is defined in `dbxout.c' and
putting a line number in desc is controlled by `#ifdef
WINNING_GDB', which defaults to false). GDB supposedly uses this
information if you say `list var'. In reality, var can
be a variable defined in the program and GDB says `function
var not defined'.
N_LSYM
stab type. Types defined at procedure scope are emited after the
N_RBRAC of the preceding function and before the code of the
procedure in which they are defined. This is exactly the same as
types defined in the source file between the two procedure bodies.
GDB overcompensates by placing all types in block #1, the block for
symbols of file scope. This is true for default, `-ansi' and
`-traditional' compiler options. (Bugs gcc/1063, gdb/1066.)
N_RBRAC or the
next N_FUN? (I believe its the first.)
N_LSYM. Sun doc talks about using N_GSYM too.
Many object file formats allow tools to create object files with custom sections containing any arbitrary data. For any such object file format, stabs can be embedded in special sections. This is how stabs are used with ELF and SOM, and aside from ECOFF and XCOFF, is how stabs are used with COFF.
The assembler creates two custom sections, a section named .stab
which contains an array of fixed length structures, one struct per stab,
and a section named .stabstr containing all the variable length
strings that are referenced by stabs in the .stab section. The
byte order of the stabs binary data depends on the object file format.
For ELF, it matches the byte order of the ELF file itself, as determined
from the EI_DATA field in the e_ident member of the ELF
header. For SOM, it is always big-endian (is this true??? FIXME). For
COFF, it matches the byte order of the COFF headers. The meaning of the
fields is the same as for a.out (see section Symbol Table Format), except
that the n_strx field is relative to the strings for the current
compilation unit (which can be found using the synthetic N_UNDF stab
described below), rather than the entire string table.
The first stab in the .stab section for each compilation unit is
synthetic, generated entirely by the assembler, with no corresponding
.stab directive as input to the assembler. This stab contains
the following fields:
n_strx
.stabstr section to the source filename.
n_type
N_UNDF.
n_other
n_desc field.
n_desc
n_value
The .stabstr section always starts with a null byte (so that string
offsets of zero reference a null string), followed by random length strings,
each of which is null byte terminated.
The ELF section header for the .stab section has its
sh_link member set to the section number of the .stabstr
section, and the .stabstr section has its ELF section
header sh_type member set to SHT_STRTAB to mark it as a
string table. SOM and COFF have no way of linking the sections together
or marking them as string tables.
For COFF, the .stab and .stabstr sections are simply
concatenated by the linker. GDB then uses the n_desc fields to
figure out the extent of the original sections. Similarly, the
n_value fields of the header symbols are added together in order
to get the actual position of the strings in a desired .stabstr
section. Although this design obviates any need for the linker to relocate
or otherwise manipulate .stab and .stabstr sections, it also
requires some care to ensure that the offsets are calculated correctly.
For instance, if the linker were to pad in between the .stabstr
sections before concatenating, then the offsets to strings in the middle
of the executable's .stabstr section would be wrong.
This section describes some Sun hacks for Stabs in ELF; it does not apply to COFF or SOM.
To keep linking fast, you don't want the linker to have to relocate very
many stabs. Making sure this is done for N_SLINE,
N_RBRAC, and N_LBRAC stabs is the most important thing
(see the descriptions of those stabs for more information). But Sun's
stabs in ELF has taken this further, to make all addresses in the
n_value field (functions and static variables) relative to the
source file. For the N_SO symbol itself, Sun simply omits the
address. To find the address of each section corresponding to a given
source file, the compiler puts out symbols giving the address of each
section for a given source file. Since these are ELF (not stab)
symbols, the linker relocates them correctly without having to touch the
stabs section. They are named Bbss.bss for the bss section,
Ddata.data for the data section, and Drodata.rodata for
the rodata section. For the text section, there is no such symbol (but
there should be, see below). For an example of how these symbols work,
See section Transformations of Stabs in separate sections. GCC does not provide these symbols;
it instead relies on the stabs getting relocated. Thus addresses which
would normally be relative to Bbss.bss, etc., are already
relocated. The Sun linker provided with Solaris 2.2 and earlier
relocates stabs using normal ELF relocation information, as it would do
for any section. Sun has been threatening to kludge their linker to not
do this (to speed up linking), even though the correct way to avoid
having the linker do these relocations is to have the compiler no longer
output relocatable values. Last I heard they had been talked out of the
linker kludge. See Sun point patch 101052-01 and Sun bug 1142109. With
the Sun compiler this affects `S' symbol descriptor stabs
(see section Static Variables) and functions (see section Procedures). In the latter
case, to adopt the clean solution (making the value of the stab relative
to the start of the compilation unit), it would be necessary to invent a
Ttext.text symbol, analogous to the Bbss.bss, etc.,
symbols. I recommend this rather than using a zero value and getting
the address from the ELF symbols.
Finding the correct Bbss.bss, etc., symbol is difficult, because
the linker simply concatenates the .stab sections from each
`.o' file without including any information about which part of a
.stab section comes from which `.o' file. The way GDB does
this is to look for an ELF STT_FILE symbol which has the same
name as the last component of the file name from the N_SO symbol
in the stabs (for example, if the file name is `../../gdb/main.c',
it looks for an ELF STT_FILE symbol named main.c). This
loses if different files have the same name (they could be in different
directories, a library could have been copied from one system to
another, etc.). It would be much cleaner to have the Bbss.bss
symbols in the stabs themselves. Having the linker relocate them there
is no more work than having the linker relocate ELF symbols, and it
solves the problem of having to associate the ELF and stab symbols.
However, no one has yet designed or implemented such a scheme.
This document was generated on 16 September 1996 using the texi2html translator version 1.51.