STABS

This document describes the stabs debugging format.

1. Overview of Stabs		Overview of stabs
2. Encoding the Structure of the Program		Encoding of the structure of the program
3. Constants
4. Variables
5. Defining Types		Type definitions
6. Symbol Information in Symbol Tables		Symbol information in symbol tables
7. GNU C++ Stabs		Stabs specific to C++
A. Table of Stab Types		Symbol types in a.out files
B. Table of Symbol Descriptors		Table of symbol descriptors
C. Table of Type Descriptors		Table of type descriptors
D. Expanded Reference by Stab Type		Reference information by stab type
E. Questions and Anomalies		Questions and anomolies
F. Using Stabs in Their Own Sections		In some object file formats, stabs are in sections.
Symbol Types Index		Index of symbolic stab symbol type names.

1. Overview of Stabs

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 B. Table of Symbol Descriptors) and type descriptors (see section C. 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.

1.1 Overview of Debugging Information Flow		Overview of debugging information flow
1.2 Overview of Stab Format		Overview of stab format
1.3 The String Field		The string field
1.4 A Simple Example in C Source		A simple example in C source
1.5 The Simple Example at the Assembly Level		The simple example at the assembly level

1.1 Overview of Debugging Information Flow

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.

1.2 Overview of Stab Format

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 6. 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 A. Table of Stab Types, for a list in numeric order of the valid type field values for stab directives.

1.3 The String Field

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 7.2 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 B. Table of Symbol Descriptors. The `c' symbol descriptor is an exception in that it is not followed by type information. See section 3. 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 C. 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, 5. Defining Types.

A type-number is often a single number. The GNU and Sun tools additionally permit a type-number to be a pair (file-number,filetype-number) (the parentheses appear in the string, and serve to distinguish the two cases). The file-number is a number starting with 1 which is incremented for each seperate source file in the compilation (e.g., in C, each header file gets a different number). The filetype-number is a number starting with 1 which is incremented for each new type defined in the file. (Separating the file number and the type number permits the N_BINCL optimization to succeed more often; see 2.3 Names of Include Files).

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 7. 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

boundary is an integer specifying the alignment. I assume it applies to all variables of this type.

pinteger

Pointer class (for checking). Not sure what this means, or how integer is interpreted.

P

Indicate this is a packed type, meaning that structure fields or array elements are placed more closely in memory, to save memory at the expense of speed.

ssize

Size in bits of a variable of this type. This is fully supported by GDB 4.11 and later.

S

Indicate that this type is a string instead of an array of characters, or a bitstring instead of a set. It doesn't change the layout of the data being represented, but does enable the debugger to know which type it is.

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.

1.4 A Simple Example in C Source

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.

1.5 The Simple Example at the Assembly Level

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

2. Encoding the Structure of the Program

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.

2.1 Main Program		Indicate what the main program is
2.2 Paths and Names of the Source Files		The path and name of the source file
2.3 Names of Include Files		Names of include files
2.4 Line Numbers
2.5 Procedures
2.6 Nested Procedures
2.7 Block Structure
2.8 Alternate Entry Points		Entering procedures except at the beginning.

2.1 Main Program

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.

2.2 Paths and Names of the Source Files

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).

2.3 Names of Include Files

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 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 GNU linker and the Sun (both SunOS4 and Solaris) linker are the only ones which supports this feature.

A linker which supports this feature will set the value of a N_BINCL symbol to the total of all the characters in the stabs strings included in the header file, omitting any file numbers. The value of an N_EXCL symbol is the same as the value of the N_BINCL symbol it replaces. This information can be used to match up N_EXCL and N_BINCL symbols which have the same filename. 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.

2.4 Line Numbers

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 F. 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 2.3 Names of Include Files).

2.5 Procedures

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 F.2 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 4.7 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 4.7.1 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.

If functions can appear in different sections, then the debugger may not be able to find the end of a function. Recent versions of GCC will mark the end of a function with an N_FUN symbol with an empty string for the name. The value is the address of the end of the current function. Without such a symbol, there is no indication of the address of the end of a function, and you must assume that it ended at the starting address of the next function or at the end of the text section for the program.

2.6 Nested Procedures

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

2.7 Block Structure

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 F. 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.

2.8 Alternate Entry Points

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.

3. Constants

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

Boolean constant. value is a numeric value; I assume it is 0 for false or 1 for true.

c value

Character constant. value is the numeric value of the constant.

e type-information , value

Constant whose value can be represented as integral. type-information is the type of the constant, as it would appear after a symbol descriptor (see section 1.3 The String Field). value is the numeric value of the constant. GDB 4.9 does not actually get the right value if value does not fit in a host 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

Integer constant. value is the numeric value. The type is some sort of generic integer type (for GDB, a host int); to specify the type explicitly, use `e' instead.

r value

Real constant. value is the real value, which can be `INF' (optionally preceded by a sign) for infinity, `QNAN' for a quiet NaN (not-a-number), or `SNAN' for a signalling NaN. If it is a normal number the format is that accepted by the C library function atof.

s string

String constant. string is a string enclosed in either `'' (in which case `'' characters within the string are represented as `\'' or `"' (in which case `"' characters within the string are represented as `\"').

S type-information , elements , bits , pattern

Set constant. type-information is the type of the constant, as it would appear after a symbol descriptor (see section 1.3 The String Field). elements is the number of elements in the set (does this means how many bits of pattern are actually used, which would be redundant with the type, or perhaps the number of bits set in pattern? I don't get it), bits is the number of bits in the constant (meaning it specifies the length of pattern, I think), and pattern is a hexadecimal representation of the set. AIX documentation refers to a limit of 32 bytes, but I see no reason why this limit should exist. This form could probably be used for arbitrary constants, not just sets; the only catch is that pattern should be understood to be target, not host, byte order and format.

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 `;'.

4. Variables

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.

4.1 Automatic Variables Allocated on the Stack		Variables allocated on the stack.
4.2 Global Variables		Variables used by more than one source file.
4.3 Register Variables		Variables in registers.
4.4 Common Blocks		Variables statically allocated together.
4.5 Static Variables		Variables local to one source file.
4.6 Fortran Based Variables		Fortran pointer based variables.
4.7 Parameters		Variables for arguments to functions.

4.1 Automatic Variables Allocated on the Stack

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 4.3 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 2.7 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 2.5 Procedures for more information on the N_FUN stab, and 2.7 Block Structure for more information on the N_LBRAC and N_RBRAC stabs.

4.2 Global Variables

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 1.3 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.

4.3 Register Variables

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.

4.4 Common Blocks

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.

4.5 Static Variables

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 F.2 Having the Linker Relocate Stabs in ELF.

4.6 Fortran Based Variables

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 4.5 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.

4.7 Parameters

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)

4.7.1 Passing Parameters in Registers
4.7.2 Storing Parameters as Local Variables
4.7.3 Passing Parameters by Reference
4.7.4 Passing Conformant Array Parameters

4.7.1 Passing Parameters in Registers

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 4.7.3 Passing Parameters by Reference).

4.7.2 Storing Parameters as Local Variables

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 is N_PSYM, here 13 is double .stabs "f:r12",64,0,3,0 # 64 is N_RSYM, here 12 is float

In both stabs 3 is the line number where f is declared (see section 2.4 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).

4.7.3 Passing Parameters by Reference

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.

4.7.4 Passing Conformant Array Parameters

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.

5. Defining Types

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.

5.1 Builtin Types		Integers, floating point, void, etc.
5.2 Miscellaneous Types		Pointers, sets, files, etc.
5.3 Cross-References to Other Types		Referring to a type not yet defined.
5.4 Subrange Types		A type with a specific range.
5.5 Array Types		An aggregate type of same-typed elements.
5.6 Strings		Like an array but also has a length.
5.7 Enumerations		Like an integer but the values have names.
5.8 Structures		An aggregate type of different-typed elements.
5.9 Giving a Type a Name		Giving a type a name.
5.10 Unions		Different types sharing storage.
5.11 Function Types

5.1 Builtin Types

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.

5.1.1 Traditional Builtin Types		Put on your seatbelts and prepare for kludgery
5.1.2 Defining Builtin Types Using Builtin Type Descriptors		Builtin types with special type descriptors
5.1.3 Negative Type Numbers		Builtin types using negative type numbers

5.1.1 Traditional Builtin Types

This is the traditional, convoluted method for defining builtin types. There are several classes of such type definitions: integer, floating point, and void.

5.1.1.1 Traditional Integer Types
5.1.1.2 Traditional Other Types

5.1.1.1 Traditional Integer Types

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.

5.1.1.2 Traditional Other Types

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.

5.1.2 Defining Builtin Types Using Builtin Type Descriptors

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 ;

Define an integral type. signed is `u' for unsigned or `s' for signed. char-flag is `c' which indicates this is a character type, or is omitted. I assume this is to distinguish an integral type from a character type of the same size, for example it might make sense to set it for the C type 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 5.2 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

Documented by AIX to define a wide character type, but their compiler actually uses negative type numbers (see section 5.1.3 Negative Type Numbers).

R fp-type ; bytes ;

Define a floating point type. fp-type has one of the following values:

1 (NF_SINGLE)

IEEE 32-bit (single precision) floating point format.

2 (NF_DOUBLE)

IEEE 64-bit (double precision) floating point format.

3 (NF_COMPLEX)

4 (NF_COMPLEX16)

5 (NF_COMPLEX32)

These are for complex numbers. A comment in the GDB source describes them as Fortran complex, double complex, and complex*16, respectively, but what does that mean? (i.e., Single precision? Double precison?).

6 (NF_LDOUBLE)

Long double. This should probably only be used for Sun format 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).

bytes is the number of bytes occupied by the type. This allows a debugger to perform some operations with the type even if it doesn't understand fp-type.

g type-information ; nbits

Documented by AIX to define a floating type, but their compiler actually uses negative type numbers (see section 5.1.3 Negative Type Numbers).

c type-information ; nbits

Documented by AIX to define a complex type, but their compiler actually uses negative type numbers (see section 5.1.3 Negative Type Numbers).

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.

5.1.3 Negative Type Numbers

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 1.3 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 5.6 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.

5.2 Miscellaneous Types

b type-information ; bytes

Pascal space type. This is documented by IBM; what does it mean?

This use of the `b' type descriptor can be distinguished from its use for builtin integral types (see section 5.1.2 Defining Builtin Types Using Builtin Type Descriptors) because the character following the type descriptor is always a digit, `(', or `-'.

B type-information

A volatile-qualified version of type-information. This is a Sun extension. References and stores to a variable with a volatile-qualified type must not be optimized or cached; they must occur as the user specifies them.

d type-information

File of type type-information. As far as I know this is only used by Pascal.

k type-information

A const-qualified version of type-information. This is a Sun extension. A variable with a const-qualified type cannot be modified.

M type-information ; length

Multiple instance type. The type seems to composed of length repetitions of type-information, for example character*3 is represented by `M-2;3', where `-2' is a reference to a character type (see section 5.1.3 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 5.4 Subrange Types.

S type-information

Pascal set type. type-information must be a small type such as an enumeration or a subrange, and the type is a bitmask whose length is specified by the number of elements in type-information.

In CHILL, if it is a bitstring instead of a set, also use the `S' type attribute (see section 1.3 The String Field).

* type-information

Pointer to type-information.

5.3 Cross-References to Other Types

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 7.2 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 5.9 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 1.3 The String Field), or what. The symbol ends with `;'.

5.4 Subrange Types

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 5.5 Array Types).

Instead of an integer, either bound can be one of the following:

A offset

The bound is passed by reference on the stack at offset offset from the argument list. See section 4.7 Parameters, for more information on such offsets.

T offset

The bound is passed by value on the stack at offset offset from the argument list.

a register-number

The bound is pased by reference in register number register-number.

t register-number

The bound is passed by value in register number register-number.

J

There is no bound.

Subranges are also used for builtin types; see 5.1.1 Traditional Builtin Types.

5.5 Array 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 5.4 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