doc/32bit.src

\C{32bit} Writing 32-bit Code (Unix, Win32, DJGPP)

This chapter attempts to cover some of the common issues involved
when writing 32-bit code, to run under \i{Win32} or Unix, or to be
linked with C code generated by a Unix-style C compiler such as
\i{DJGPP}. It covers how to write assembly code to interface with
32-bit C routines, and how to write position-independent code for
shared libraries.

Almost all 32-bit code, and in particular all code running under
\c{Win32}, \c{DJGPP} or any of the PC Unix variants, runs in \I{flat
memory model}\e{flat} memory model. This means that the segment registers
and paging have already been set up to give you the same 32-bit 4Gb
address space no matter what segment you work relative to, and that
you should ignore all segment registers completely. When writing
flat-model application code, you never need to use a segment
override or modify any segment register, and the code-section
addresses you pass to \c{CALL} and \c{JMP} live in the same address
space as the data-section addresses you access your variables by and
the stack-section addresses you access local variables and procedure
parameters by. Every address is 32 bits long and contains only an
offset part.


\H{32c} Interfacing to 32-bit C Programs

A lot of the discussion in \k{16c}, about interfacing to 16-bit C
programs, still applies when working in 32 bits. The absence of
memory models or segmentation worries simplifies things a lot.


\S{32cunder} External Symbol Names

Most 32-bit C compilers share the convention used by 16-bit
compilers, that the names of all global symbols (functions or data)
they define are formed by prefixing an underscore to the name as it
appears in the C program. However, not all of them do: the \c{ELF}
specification states that C symbols do \e{not} have a leading
underscore on their assembly-language names.

The older Linux \c{a.out} C compiler, all \c{Win32} compilers,
\c{DJGPP}, and \c{NetBSD} and \c{FreeBSD}, all use the leading
underscore; for these compilers, the macros \c{cextern} and
\c{cglobal}, as given in \k{16cunder}, will still work. For \c{ELF},
though, the leading underscore should not be used.

See also \k{opt-pfix}.

\S{32cfunc} Function Definitions and Function Calls

\I{functions, C calling convention}The \i{C calling convention}
in 32-bit programs is as follows. In the following description,
the words \e{caller} and \e{callee} are used to denote
the function doing the calling and the function which gets called.

\b The caller pushes the function's parameters on the stack, one
after another, in reverse order (right to left, so that the first
argument specified to the function is pushed last).

\b The caller then executes a near \c{CALL} instruction to pass
control to the callee.

\b The callee receives control, and typically (although this is not
actually necessary, in functions which do not need to access their
parameters) starts by saving the value of \c{ESP} in \c{EBP} so as
to be able to use \c{EBP} as a base pointer to find its parameters
on the stack. However, the caller was probably doing this too, so
part of the calling convention states that \c{EBP} must be preserved
by any C function. Hence the callee, if it is going to set up
\c{EBP} as a \i{frame pointer}, must push the previous value first.

\b The callee may then access its parameters relative to \c{EBP}.
The doubleword at \c{[EBP]} holds the previous value of \c{EBP} as
it was pushed; the next doubleword, at \c{[EBP+4]}, holds the return
address, pushed implicitly by \c{CALL}. The parameters start after
that, at \c{[EBP+8]}. The leftmost parameter of the function, since
it was pushed last, is accessible at this offset from \c{EBP}; the
others follow, at successively greater offsets. Thus, in a function
such as \c{printf} which takes a variable number of parameters, the
pushing of the parameters in reverse order means that the function
knows where to find its first parameter, which tells it the number
and type of the remaining ones.

\b The callee may also wish to decrease \c{ESP} further, so as to
allocate space on the stack for local variables, which will then be
accessible at negative offsets from \c{EBP}.

\b The callee, if it wishes to return a value to the caller, should
leave the value in \c{AL}, \c{AX} or \c{EAX} depending on the size
of the value. Floating-point results are typically returned in
\c{ST0}.

\b Once the callee has finished processing, it restores \c{ESP} from
\c{EBP} if it had allocated local stack space, then pops the previous
value of \c{EBP}, and returns via \c{RET} (equivalently, \c{RETN}).

\b When the caller regains control from the callee, the function
parameters are still on the stack, so it typically adds an immediate
constant to \c{ESP} to remove them (instead of executing a number of
slow \c{POP} instructions). Thus, if a function is accidentally
called with the wrong number of parameters due to a prototype
mismatch, the stack will still be returned to a sensible state since
the caller, which \e{knows} how many parameters it pushed, does the
removing.

There is an alternative calling convention used by Win32 programs
for Windows API calls, and also for functions called \e{by} the
Windows API such as window procedures: they follow what Microsoft
calls the \c{__stdcall} convention. This is slightly closer to the
Pascal convention, in that the callee clears the stack by passing a
parameter to the \c{RET} instruction. However, the parameters are
still pushed in right-to-left order.

Thus, you would define a function in C style in the following way:

\c global  _myfunc
\c
\c _myfunc:
\c         push    ebp
\c         mov     ebp,esp
\c         sub     esp,0x40        ; 64 bytes of local stack space
\c         mov     ebx,[ebp+8]     ; first parameter to function
\c
\c         ; some more code
\c
\c         leave                   ; mov esp,ebp / pop ebp
\c         ret

At the other end of the process, to call a C function from your
assembly code, you would do something like this:

\c extern  _printf
\c
\c         ; and then, further down...
\c
\c         push    dword [myint]   ; one of my integer variables
\c         push    dword mystring  ; pointer into my data segment
\c         call    _printf
\c         add     esp,byte 8      ; `byte' saves space
\c
\c         ; then those data items...
\c
\c segment _DATA
\c
\c myint       dd   1234
\c mystring    db   'This number -> %d <- should be 1234',10,0

This piece of code is the assembly equivalent of the C code

\c     int myint = 1234;
\c     printf("This number -> %d <- should be 1234\n", myint);


\S{32cdata} Accessing Data Items

To get at the contents of C variables, or to declare variables which
C can access, you need only declare the names as \c{GLOBAL} or
\c{EXTERN}. (Again, the names require leading underscores, as stated
in \k{32cunder}.) Thus, a C variable declared as \c{int i} can be
accessed from assembler as

\c           extern _i
\c           mov eax,[_i]

And to declare your own integer variable which C programs can access
as \c{extern int j}, you do this (making sure you are assembling in
the \c{_DATA} segment, if necessary):

\c           global _j
\c _j        dd 0

To access a C array, you need to know the size of the components of
the array. For example, \c{int} variables are four bytes long, so if
a C program declares an array as \c{int a[10]}, you can access
\c{a[3]} by coding \c{mov ax,[_a+12]}. (The byte offset 12 is obtained
by multiplying the desired array index, 3, by the size of the array
element, 4.) The sizes of the C base types in 32-bit compilers are:
1 for \c{char}, 2 for \c{short}, 4 for \c{int}, \c{long} and
\c{float}, and 8 for \c{double}. Pointers, being 32-bit addresses,
are also 4 bytes long.

To access a C \i{data structure}, you need to know the offset from
the base of the structure to the field you are interested in. You
can either do this by converting the C structure definition into a
NASM structure definition (using \c{STRUC}), or by calculating the
one offset and using just that.

To do either of these, you should read your C compiler's manual to
find out how it organizes data structures. NASM gives no special
alignment to structure members in its own \i\c{STRUC} macro, so you
have to specify alignment yourself if the C compiler generates it.
Typically, you might find that a structure like

\c struct {
\c     char c;
\c     int i;
\c } foo;

might be eight bytes long rather than five, since the \c{int} field
would be aligned to a four-byte boundary. However, this sort of
feature is sometimes a configurable option in the C compiler, either
using command-line options or \c{#pragma} lines, so you have to find
out how your own compiler does it.


\S{32cmacro} \i\c{c32.mac}: Helper Macros for the 32-bit C Interface

Included in the NASM archives, in the \I{misc directory}\c{misc}
directory, is a file \c{c32.mac} of macros. It defines three macros:
\i\c{proc}, \i\c{arg} and \i\c{endproc}. These are intended to be
used for C-style procedure definitions, and they automate a lot of
the work involved in keeping track of the calling convention.

An example of an assembly function using the macro set is given
here:

\c proc    _proc32
\c
\c %$i     arg
\c %$j     arg
\c         mov     eax,[ebp + %$i]
\c         mov     ebx,[ebp + %$j]
\c         add     eax,[ebx]
\c
\c endproc

This defines \c{_proc32} to be a procedure taking two arguments, the
first (\c{i}) an integer and the second (\c{j}) a pointer to an
integer. It returns \c{i + *j}.

Note that the \c{arg} macro has an \c{EQU} as the first line of its
expansion, and since the label before the macro call gets prepended
to the first line of the expanded macro, the \c{EQU} works, defining
\c{%$i} to be an offset from \c{BP}. A context-local variable is
used, local to the context pushed by the \c{proc} macro and popped
by the \c{endproc} macro, so that the same argument name can be used
in later procedures. Of course, you don't \e{have} to do that.

\c{arg} can take an optional parameter, giving the size of the
argument. If no size is given, 4 is assumed, since it is likely that
many function parameters will be of type \c{int} or pointers.


\H{picdll} Writing NetBSD/FreeBSD/OpenBSD and Linux/ELF \i{Shared
Libraries}

\I{Linux}\i{ELF} replaced the older \c{a.out} object file format
under Linux because it contains support for \i{position-independent
code} (\i{PIC}), which makes writing shared libraries much
easier. NASM supports the \c{ELF} position-independent code features,
so you can write Linux \c{ELF} shared libraries in NASM.

\i{NetBSD}, and its close cousins \i{FreeBSD} and \i{OpenBSD}, take
a different approach by hacking PIC support into the \c{a.out}
format. NASM supports this as the \i\c{aoutb} output format, so you
can write \i{BSD} shared libraries in NASM too.

The operating system loads a PIC shared library by memory-mapping
the library file at an arbitrarily chosen point in the address space
of the running process. The contents of the library's code section
must therefore not depend on where it is loaded in memory.

Therefore, you cannot get at your variables by writing code like
this:

\c         mov     eax,[myvar]             ; WRONG

Instead, the linker provides an area of memory called the
\i\e{global offset table}, or \i{GOT}; the GOT is situated at a
constant distance from your library's code, so if you can find out
where your library is loaded (which is typically done using a
\c{CALL} and \c{POP} combination), you can obtain the address of the
GOT, and you can then load the addresses of your variables out of
linker-generated entries in the GOT.

The \e{data} section of a PIC shared library does not have these
restrictions: since the data section is writable, it has to be
copied into memory anyway rather than just paged in from the library
file, so as long as it's being copied it can be relocated too. So
you can put ordinary types of relocation in the data section without
too much worry (but see \k{picglobal} for a caveat).


\S{picgot} Obtaining the Address of the GOT

Each code module in your shared library should define the GOT as an
external symbol:

\c extern  _GLOBAL_OFFSET_TABLE_   ; in ELF
\c extern  __GLOBAL_OFFSET_TABLE_  ; in BSD a.out

At the beginning of any function in your shared library which plans
to access your data or BSS sections, you must first calculate the
address of the GOT. This is typically done by writing the function
in this form:

\c func:   push    ebp
\c         mov     ebp,esp
\c         push    ebx
\c         call    .get_GOT
\c .get_GOT:
\c         pop     ebx
\c         add     ebx,_GLOBAL_OFFSET_TABLE_+$$-.get_GOT wrt ..gotpc
\c
\c         ; the function body comes here
\c
\c         mov     ebx,[ebp-4]
\c         mov     esp,ebp
\c         pop     ebp
\c         ret

(For BSD, again, the symbol \c{_GLOBAL_OFFSET_TABLE} requires a
second leading underscore.)

The first two lines of this function are simply the standard C
prologue to set up a stack frame, and the last three lines are
standard C function epilogue. The third line, and the fourth to last
line, save and restore the \c{EBX} register, because PIC shared
libraries use this register to store the address of the GOT.

The interesting bit is the \c{CALL} instruction and the following
two lines. The \c{CALL} and \c{POP} combination obtains the address
of the label \c{.get_GOT}, without having to know in advance where
the program was loaded (since the \c{CALL} instruction is encoded
relative to the current position). The \c{ADD} instruction makes use
of one of the special PIC relocation types: \i{GOTPC relocation}.
With the \i\c{WRT ..gotpc} qualifier specified, the symbol
referenced (here \c{_GLOBAL_OFFSET_TABLE_}, the special symbol
assigned to the GOT) is given as an offset from the beginning of the
section. (Actually, \c{ELF} encodes it as the offset from the operand
field of the \c{ADD} instruction, but NASM simplifies this
deliberately, so you do things the same way for both \c{ELF} and
\c{BSD}.) So the instruction then \e{adds} the beginning of the section,
to get the real address of the GOT, and subtracts the value of
\c{.get_GOT} which it knows is in \c{EBX}. Therefore, by the time
that instruction has finished, \c{EBX} contains the address of the GOT.

If you didn't follow that, don't worry: it's never necessary to
obtain the address of the GOT by any other means, so you can put
those three instructions into a macro and safely ignore them:

\c %macro  get_GOT 0
\c
\c         call    %%getgot
\c   %%getgot:
\c         pop     ebx
\c         add     ebx,_GLOBAL_OFFSET_TABLE_+$$-%%getgot wrt ..gotpc
\c
\c %endmacro

\S{piclocal} Finding Your Local Data Items

Having got the GOT, you can then use it to obtain the addresses of
your data items. Most variables will reside in the sections you have
declared; they can be accessed using the \I{GOTOFF
relocation}\c{..gotoff} special \I\c{WRT ..gotoff}\c{WRT} type. The
way this works is like this:

\c         lea     eax,[ebx+myvar wrt ..gotoff]

The expression \c{myvar wrt ..gotoff} is calculated, when the shared
library is linked, to be the offset to the local variable \c{myvar}
from the beginning of the GOT. Therefore, adding it to \c{EBX} as
above will place the real address of \c{myvar} in \c{EAX}.

If you declare variables as \c{GLOBAL} without specifying a size for
them, they are shared between code modules in the library, but do
not get exported from the library to the program that loaded it.
They will still be in your ordinary data and BSS sections, so you
can access them in the same way as local variables, using the above
\c{..gotoff} mechanism.

Note that due to a peculiarity of the way BSD \c{a.out} format
handles this relocation type, there must be at least one non-local
symbol in the same section as the address you're trying to access.


\S{picextern} Finding External and Common Data Items

If your library needs to get at an external variable (external to
the \e{library}, not just to one of the modules within it), you must
use the \I{GOT relocations}\I\c{WRT ..got}\c{..got} type to get at
it. The \c{..got} type, instead of giving you the offset from the
GOT base to the variable, gives you the offset from the GOT base to
a GOT \e{entry} containing the address of the variable. The linker
will set up this GOT entry when it builds the library, and the
dynamic linker will place the correct address in it at load time. So
to obtain the address of an external variable \c{extvar} in \c{EAX},
you would code

\c         mov     eax,[ebx+extvar wrt ..got]

This loads the address of \c{extvar} out of an entry in the GOT. The
linker, when it builds the shared library, collects together every
relocation of type \c{..got}, and builds the GOT so as to ensure it
has every necessary entry present.

Common variables must also be accessed in this way.


\S{picglobal} Exporting Symbols to the Library User

If you want to export symbols to the user of the library, you have to
declare whether they are functions or data, and if they are data, you
have to give the size of the data item. This is because the dynamic
linker has to build \I{PLT}\e{procedure linkage table} entries for any
exported functions, and also moves exported data items away from the
library's data section in which they were declared.

So to export a function to users of the library, you must use

\c global  func:function           ; declare it as a function
\c
\c func:   push    ebp
\c
\c         ; etc.

And to export a data item such as an array, you would have to code

\c global  array:data array.end-array      ; give the size too
\c
\c array:  resd    128
\c .end:

Be careful: If you export a variable to the library user, by
declaring it as \c{GLOBAL} and supplying a size, the variable will
end up living in the data section of the main program, rather than
in your library's data section, where you declared it. So you will
have to access your own global variable with the \c{..got} mechanism
rather than \c{..gotoff}, as if it were external (which,
effectively, it has become).

Equally, if you need to store the address of an exported global in
one of your data sections, you can't do it by means of the standard
sort of code:

\c dataptr:        dd      global_data_item        ; WRONG

NASM will interpret this code as an ordinary relocation, in which
\c{global_data_item} is merely an offset from the beginning of the
\c{.data} section (or whatever); so this reference will end up
pointing at your data section instead of at the exported global
which resides elsewhere.

Instead of the above code, then, you must write

\c dataptr:        dd      global_data_item wrt ..sym

which makes use of the special \c{WRT} type \I\c{WRT ..sym}\c{..sym}
to instruct NASM to search the symbol table for a particular symbol
at that address, rather than just relocating by section base.

Either method will work for functions: referring to one of your
functions by means of

\c funcptr:        dd      my_function

will give the user the address of the code you wrote, whereas

\c funcptr:        dd      my_function wrt ..sym

will give the address of the procedure linkage table for the
function, which is where the calling program will \e{believe} the
function lives. Either address is a valid way to call the function.


\S{picproc} Calling Procedures Outside the Library

Calling procedures outside your shared library has to be done by means
of a \I{PLT relocations}\e{procedure linkage table}, or \i{PLT}. The
PLT is placed at a known offset from where the library is loaded, so
the library code can make calls to the PLT in a position-independent
way. Within the PLT there is code to jump to offsets contained in the
GOT, so function calls to other shared libraries or to routines in the
main program can be transparently passed off to their real
destinations.

To call an external routine, you must use another special PIC
relocation type, \I{PLT relocations}\i\c{WRT ..plt}. This is much
easier than the GOT-based ones: you simply replace calls such as
\c{CALL printf} with the PLT-relative version \c{CALL printf WRT
..plt}.


\S{link} Generating the Library File

Having written some code modules and assembled them to \c{.o} files,
you then generate your shared library with a command such as

\c ld -shared -o library.so module1.o module2.o       # for ELF
\c ld -Bshareable -o library.so module1.o module2.o   # for BSD

For ELF, if your shared library is going to reside in system
directories such as \c{/usr/lib} or \c{/lib}, it is usually worth
using the \i\c{-soname} flag to the linker, to store the final
library file name, with a version number, into the library:

\c ld -shared -soname library.so.1 -o library.so.1.2 *.o

You would then copy \c{library.so.1.2} into the library directory,
and create \c{library.so.1} as a symbolic link to it.