path: root/en_US.ISO8859-1/books/developers-handbook/x86/chapter.sgml

<!--
     The FreeBSD Documentation Project

     This file is automatically generated.  Please do not make commits
     to this file.  Updates should be sent to the author :

     G. Adam Stanislav (adam@whizkidtech.net)

     This chapter is an exception to our general rule, and the author
     retains the copyright.  Among other things, this means that this
     chapter should not be included in any printed version of the
     Developer's Handbook without Adam's explicit permission.

     Eventually we will have to replace this chapter or convince the
     author to assign us the copyright.  For now, it is valuable
     content so it should stay.

     $FreeBSD: doc/en_US.ISO8859-1/books/developers-handbook/x86/chapter.sgml,v 1.4 2001/06/23 06:56:59 dd Exp $
-->

<chapter id="x86">

<title>x86 Assembly Language Programming</title>
<para>
<emphasis>
This chapter was written by G. Adam Stanislav.
<ulink url="http://www.whizkidtech.net/">Whiz Kid Technomagic</ulink>
</emphasis></para>



<sect1 id="x86-intro">
<title>Synopsis</title>

<para>
Assembly language programing under Unix is highly undocumented. It
is generally assumed that no one would ever want to use it because
various Unix systems run on different microprocessors, so everything
should be written in C for portability.
</para>

<para>
In reality, C portability is quite a myth. Even C programs need
to be modified when ported from one Unix to another, regardless of
what processor each runs on. Typically, such a program is full
of conditional statements depending on the system it is
compiled for.
</para>

<para>
Even if we believe that all of Unix software should be written in C,
or some other high-level language, we still need assembly language
programmers: Who else would write the section of C library
that accesses the kernel?
</para>

<para>
In this chapter I will attempt to show you
how you can use assembly language writing
Unix programs, specifically under FreeBSD.
</para>

<para>
This chapter does not explain the basics of assembly language.
There are enough resources about that (for a complete
online course in assembly language, see Randall Hyde's
<ulink url="http://webster.cs.ucr.edu/">Art
of Assembly Language</ulink>; or if you prefer
a printed book, take a look at Jeff Duntemann's
<ulink url="http://www.int80h.org/cgi-bin/isbn?isbn=0471375233">Assembly
Language Step-by-Step</ulink>). However,
once the chapter is finished, any assembly language programmer
will be able to write programs for FreeBSD
quickly and efficiently.
</para>

<para>
Copyright &copy; 2000-2001 G. Adam Stanislav. All rights reserved.
</para>

</sect1>

<sect1 id="x86-the-tools">
<title>The Tools</title>

<sect2 id="x86-the-assembler">
<title>The Assembler</title>

<para>
The most important tool for assembly language programming is the
assembler, the software that converts assembly language code
into machine language.
</para>

<para>
Two very different assemblers are available for FreeBSD. One is
<citerefentry><refentrytitle>as</refentrytitle><manvolnum>1</manvolnum></citerefentry>,
which uses the traditional Unix assembly language syntax. It
comes with the system.
</para>

<para>
The other is <application>/usr/ports/devel/nasm</application>.
It uses the Intel syntax. Its main advantage is that it
can assemble code for many operating systems. It needs
to be installed separately, but is completely free.
</para>

<para>
This chapter uses <application>nasm</application>
syntax because most assembly language programmers
coming to FreeBSD from other operating systems
will find it easier to understand. And, because,
quite frankly, that is what I am used to.
</para>

</sect2>

<sect2 id="x86-the-linker">
<title>The Linker</title>

<para>
The output of the assembler, like that of any
compiler, needs to be linked to form an executable file.
</para>

<para>
The standard
<citerefentry><refentrytitle>ld</refentrytitle><manvolnum>1</manvolnum></citerefentry>
linker comes with FreeBSD. It works with the
code assembled with either assembler.
</para>

</sect2>
</sect1>

<sect1 id="x86-system-calls">
<title>System Calls</title>

<sect2 id="x86-default-calling-convention">
<title>Default Calling Convention</title>

<para>
By default, the FreeBSD kernel uses the C calling
convention. Further, although the kernel is accessed
using <function role="opcode">int 80h</function>,
it is assumed the program will call a function that
issues <function role="opcode">int 80h</function>, rather than
issuing <function role="opcode">int 80h</function> directly.
</para>

<para>
This convention is very convenient, and quite superior to the
Microsoft convention used by <acronym>MS DOS</acronym>.
Why? Because the Unix convention allows any program written in
any language to access the kernel.
</para>

<para>
An assembly language program can do that as well.
For example, we could open a file:
</para>

<programlisting>
kernel:
	int	80h	; Call kernel
	ret

open:
	push	dword mode
	push	dword flags
	push	dword path
	mov	eax, 5
	call	kernel
	add	esp, byte 12
	ret
</programlisting>

<para>
This is a very clean and portable way of coding. If you need to
port the code to a Unix system which uses a different interrupt,
or a different way of passing parameters, all you need to change
is the kernel procedure.
</para>

<para>
But assembly language programmers like to shave off cycles. The above example
requires a <function role="opcode">call/ret</function> combination.
We can eliminate it by
<function role="opcode">push</function>ing an extra dword:
</para>

<programlisting>
open:
	push	dword mode
	push	dword flags
	push	dword path
	mov	eax, 5
	push	eax		; Or any other dword
	int	80h
	add	esp, byte 16
</programlisting>

<para>
The <constant>5</constant> that we have placed in
<varname role="register">EAX</varname> identifies
the kernel function, in this case <function role="syscall">open</function>.
</para>

</sect2>
<sect2 id="x86-alternate-calling-convention">
<title>Alternate Calling Convention</title>
<para>
FreeBSD is an extremely flexible system. It offers other ways of
calling the kernel. For it to work, however, the system must
have Linux emulation installed.
</para>

<para>
Linux is a Unix-like system. However, its kernel uses the same
system-call convention of passing parameters in registers
<acronym>MS DOS</acronym> does. As with the Unix convention,
the function number is placed in <varname role="register">EAX</varname>.
The parameters, however, are not passed on the stack but in
<varname role="register">EBX, ECX, EDX, ESI, EDI, EBP</varname>:
</para>

<programlisting>
open:
	mov	eax, 5
	mov	ebx, path
	mov	ecx, flags
	mov	edx, mode
	int	80h
</programlisting>

<para>
This convention has a great disadvantage over
the Unix way, at least as far as assembly language programming
is concerned: Every time you make a kernel call
you must <function role="opcode">push</function> the registers, then
<function role="opcode">pop</function> them later. This makes your code
bulkier and slower. Nevertheless, FreeBSD gives
you a choice.
</para>

<para>
If you do choose the Linux convention, you must let
the system know about it. After your program is assembled and
linked, you need to brand the executable:
</para>

<screen>&prompt.user; <userinput>brandelf -f Linux <replaceable>filename</replaceable></userinput></screen>

</sect2>

<sect2 id="x86-use-geneva">
<title>Which Convention Should You Use?</title>

<para>
If you are coding specifically for FreeBSD, you should always
use the Unix convention: It is faster, you can store global
variables in registers, you do not have to brand
the executable, and you do not impose the installation of
the Linux emulation package on the target system.
</para>

<para>
If you want to create portable code that can also run
on Linux, you will probably still want to give the FreeBSD
users as efficient a code as possible. I will show you
how you can accomplish that after I have explained the basics.
</para>

</sect2>

<sect2 id="x86-call-numbers">
<title>Call Numbers</title>

<para>
To tell the kernel which system service you are calling,
place its number in <varname role="register">EAX</varname>. Of course, you need
to know what the number is.
</para>

<sect3 id="x86-the-syscalls-file">
<title>The <filename>syscalls</filename> File</title>

<para>
The numbers are listed in <filename>syscalls</filename>.
<command>locate syscalls</command> finds this file
in several different formats, all produced automatically
from <filename>syscalls.master</filename>.
</para>

<para>
You can find the master file for the default Unix calling
convention in
<filename>/usr/src/sys/kern/syscalls.master</filename>.
If you need to use the other convention implemented
in the Linux emulation mode, read
<filename>/usr/src/sys/i386/linux/syscalls.master</filename>.
</para>

<note>
<para>
Not only do FreeBSD and Linux use different calling
conventions, they sometimes use different numbers for
the same functions.
</para>
</note>

<para>
<filename>syscalls.master</filename> describes how
the call is to be made:
</para>

<programlisting>
0	STD	NOHIDE	{ int nosys(void); } syscall nosys_args int
1	STD	NOHIDE	{ void exit(int rval); } exit rexit_args void
2	STD	POSIX	{ int fork(void); }
3	STD	POSIX	{ ssize_t read(int fd, void *buf, size_t nbyte); }
4	STD	POSIX	{ ssize_t write(int fd, const void *buf, size_t nbyte); }
5	STD	POSIX	{ int open(char *path, int flags, int mode); }
6	STD	POSIX	{ int close(int fd); }
etc...
</programlisting>
<para>
It is the leftmost column that tells us the number to place in
<varname role="register">EAX</varname>.
</para>

<para>
The rightmost column tells us what parameters to
<function role="opcode">push</function>. They are <function role="opcode">push</function>ed
<emphasis>from right to left</emphasis>.
</para>

<informalexample>
<para>
For example, to <function>open</function> a file, we need
to <function role="opcode">push</function> the <varname>mode</varname> first,
then <varname>flags</varname>, then the address at which
the <varname>path</varname> is stored.
</para>
</informalexample>

</sect3>

</sect2>

</sect1>

<sect1 id="x86-return-values">
<title>Return Values</title>

<para>
A system call would not be useful most of the time
if it did not return some kind of a value: The file
descriptor of an open file, the number of bytes read
to a buffer, the system time, etc.
</para>

<para>
Additionally, the system needs to inform us if an error
occurs: A file does not exist, system resources are exhausted,
we passed an invalid parameter, etc.
</para>

<sect2 id="x86-man-pages">
<title>Man Pages</title>

<para>
The traditional place to look for information about various
system calls under Unix systems are the man pages.
FreeBSD describes its system calls in section 2, sometimes
in section 3.
</para>

<para>
For example, <citerefentry><refentrytitle>open</refentrytitle><manvolnum>2</manvolnum></citerefentry> says:
</para>

<blockquote>
<para>
If successful, <function>open()</function> returns a non-negative
integer, termed a file descriptor. It returns <varname>-1</varname> on failure,
and sets <varname>errno</varname> to indicate the error.
</para>

</blockquote>
<para>
The assembly language programmer new to Unix and FreeBSD will
immediately ask the puzzling question: Where is
<varname>errno</varname> and how do I get to it?
</para>

<note>
<para>
The information presented in the man pages applies
to C programs. The assembly language programmer needs additional
information.
</para>
</note>

</sect2>

<sect2 id="x86-where-return-values">
<title>Where Are the Return Values?</title>

<para>
Unfortunately, it depends... For most system calls it is
in <varname role="register">EAX</varname>, but not for all.
A good rule of thumb,
when working with a system call for
the first time, is to look for
the return value in <varname role="register">EAX</varname>.
If it is not there, you
need further research.
</para>

<note>
<para>
I am aware of one system call that returns the value in
<varname role="register">EDX</varname>: <function role="syscall">SYS_fork</function>. All others
I have worked with use <varname role="register">EAX</varname>.
But I have not worked with them all yet.
</para>
</note>

<tip>
<para>
If you cannot find the answer here or anywhere else,
study <application>libc</application> source code and see how it
interfaces with the kernel.
</para>
</tip>

</sect2>
<sect2 id="x86-where-errno">
<title>Where Is <varname>errno</varname>?</title>

<para>
Actually, nowhere...
</para>

<para>
<varname>errno</varname> is part of the C language, not the
Unix kernel. When accessing kernel services directly, the
error code is returned in <varname role="register">EAX</varname>,
the same register the proper
return value generally ends up in.
</para>

<para>
This makes perfect sense. If there is no error, there is
no error code. If there is an error, there is no return
value. One register can contain either.
</para>

</sect2>

<sect2 id="x86-how-to-know-error">
<title>Determining an Error Occurred</title>

<para>
When using the standard FreeBSD calling convention,
the <varname role="register">carry flag</varname> is cleared upon success,
set upon failure.
</para>

<para>
When using the Linux emulation mode, the signed
value in <varname role="register">EAX</varname> is non-negative upon success,
and contains the return value. In case of an error, the value
is negative, i.e., <varname>-errno</varname>.
</para>

</sect2>

</sect1>

<sect1 id="x86-portable-code">
<title>Creating Portable Code</title>

<para>
Portability is generally not one of the strengths of assembly language.
Yet, writing assembly language programs for different platforms is
possible, especially with <application>nasm</application>. I have written
assembly language libraries that can be assembled for such different
operating systems as Windows and FreeBSD.
</para>

<para>
It is all the more possible when you want your code to run
on two platforms which, while different, are based on
similar architectures.
</para>

<para>
For example, FreeBSD is Unix, Linux is Unix-like. I only
mentioned three differences between them (from an assembly language
programmer's perspective): The calling convention, the
function numbers, and the way of returning values.
</para>

<sect2 id="x86-deal-with-function-numbers"><title>Dealing with Function Numbers</title>

<para>
In many cases the function numbers are the same. However,
even when they are not, the problem is easy to deal with:
Instead of using numbers in your code, use constants which
you have declared differently depending on the target
architecture:
</para>

<programlisting>
%ifdef	LINUX
%define	SYS_execve	11
%else
%define	SYS_execve	59
%endif
</programlisting>
</sect2>
<sect2 id="x86-deal-with-geneva"><title>Dealing with Conventions</title>
<para>
Both, the calling convention, and the return value (the
<varname>errno</varname> problem) can be resolved with macros:
</para>

<programlisting>
%ifdef	LINUX

%macro	system	0
	call	kernel
%endmacro

align 4
kernel:
	push	ebx
	push	ecx
	push	edx
	push	esi
	push	edi
	push	ebp

	mov	ebx, [esp+32]
	mov	ecx, [esp+36]
	mov	edx, [esp+40]
	mov	esi, [esp+44]
	mov	ebp, [esp+48]
	int	80h

	pop	ebp
	pop	edi
	pop	esi
	pop	edx
	pop	ecx
	pop	ebx

	or	eax, eax
	js	.errno
	clc
	ret

.errno:
	neg	eax
	stc
	ret

%else

%macro	system	0
	int	80h
%endmacro

%endif
</programlisting>

</sect2>

<sect2 id="x86-deal-with-other-portability"><title>Dealing with Other Portability Issues</title>

<para>
The above solutions can handle most cases of writing code
portable between FreeBSD and Linux. Nevertheless, with some
kernel services the differences are deeper.
</para>

<para>
In that case, you need to write two different handlers
for those particular system calls, and use conditional
assembly. Luckily, most of your code does something other
than calling the kernel, so usually you will only need
a few such conditional sections in your code.
</para>

</sect2>

<sect2 id="x86-portable-library"><title>Using a Library</title>

<para>
You can avoid portability issues in your main code altogether
by writing a library of system calls. Create a separate library
for FreeBSD, a different one for Linux, and yet other libraries
for more operating systems.
</para>

<para>
In your library, write a separate function (or procedure, if
you prefer the traditional assembly language terminology) for each system
call. Use the C calling convention of passing parameters.
But still use <varname role="register">EAX</varname> to pass the call number in.
In that case, your FreeBSD library can be very simple, as
many seemingly different functions can be just labels to
the same code:
</para>

<programlisting>
sys.open:
sys.close:
[etc...]
	int	80h
	ret
</programlisting>

<para>
Your Linux library will require more different functions.
But even here you can group system calls using the same
number of parameters:
</para>

<programlisting>
sys.exit:
sys.close:
[etc... one-parameter functions]
	push	ebx
	mov	ebx, [esp+12]
	int	80h
	pop	ebx
	jmp	sys.return

...

sys.return:
	or	eax, eax
	js	sys.err
	clc
	ret

sys.err:
	neg	eax
	stc
	ret
</programlisting>

<para>
The library approach may seem inconvenient at first because
it requires you to produce a separate file your code depends
on. But it has many advantages: For one, you only need to
write it once and can use it for all your programs. You can
even let other assembly language programmers use it, or perhaps use
one written by someone else. But perhaps the greatest
advantage of the library is that your code can be ported
to other systems, even by other programmers, by simply
writing a new library without any changes to your code.
</para>

<para>
If you do not like the idea of having a library, you can
at least place all your system calls in a separate assembly language file
and link it with your main program. Here, again, all porters
have to do is create a new object file to link with your
main program.
</para>

</sect2>

<sect2 id="x86-portable-include">
<title>Using an Include File</title>

<para>
If you are releasing your software as (or with)
source code, you can use macros and place them
in a separate file, which you include in your
code.
</para>

<para>
Porters of your software will simply write a new
include file. No library or external object file
is necessary, yet your code is portable without any
need to edit the code.
</para>

<note>
<para>
This is the approach we will use throughout this chapter.
We will name our include file <filename>system.inc</filename>, and
add to it whenever we deal with a new system call.
</para>
</note>

<para>
We can start our <filename>system.inc</filename> by declaring the
standard file descriptors:
</para>

<programlisting>
%define	stdin	0
%define	stdout	1
%define	stderr	2
</programlisting>

<para>
Next, we create a symbolic name for each system call:
</para>

<programlisting>
%define	SYS_nosys	0
%define	SYS_exit	1
%define	SYS_fork	2
%define	SYS_read	3
%define	SYS_write	4
; [etc...]
</programlisting>

<para>
We add a short, non-global procedure with a long name,
so we do not accidentally reuse the name in our code:
</para>

<programlisting>
section	.text
align 4
access.the.bsd.kernel:
	int	80h
	ret
</programlisting>

<para>
We create a macro which takes one argument, the syscall number:
</para>

<programlisting>
%macro	system	1
	mov	eax, %1
	call	access.the.bsd.kernel
%endmacro
</programlisting>

<para>
Finally, we create macros for each syscall. These macros take
no arguments.
</para>

<programlisting>
%macro	sys.exit	0
	system	SYS_exit
%endmacro

%macro	sys.fork	0
	system	SYS_fork
%endmacro

%macro	sys.read	0
	system	SYS_read
%endmacro

%macro	sys.write	0
	system	SYS_write
%endmacro

; [etc...]
</programlisting>

<para>
Go ahead, enter it into your editor and save it as
<filename>system.inc</filename>. We will add more to it as we
discuss more syscalls.
</para>

</sect2>

</sect1>

<sect1 id="x86-first-program">
<title>Our First Program</title>

<para>
We are now ready for our first program, the mandatory
<application>Hello, World!</application>
</para>

<programlisting>
 1:	%include	'system.inc'
 2:
 3:	section	.data
 4:	hello	db	'Hello, World!', 0Ah
 5:	hbytes	equ	$-hello
 6:
 7:	section	.text
 8:	global	_start
 9:	_start:
10:	push	dword hbytes
11:	push	dword hello
12:	push	dword stdout
13:	sys.write
14:
15:	push	dword 0
16:	sys.exit
</programlisting>

<para>
Here is what it does: Line 1 includes the defines, the macros,
and the code from <filename>system.inc</filename>.
</para>

<para>
Lines 3-5 are the data: Line 3 starts the data section/segment.
Line 4 contains the string "Hello, World!" followed by a new
line (<constant>0Ah</constant>). Line 5 creates a constant that contains
the length of the string from line 4 in bytes.
</para>

<para>
Lines 7-16 contain the code. Note that FreeBSD uses the <emphasis>elf</emphasis>
file format for its executables, which requires every
program to start at the point labeled <varname>_start</varname> (or, more
precisely, the linker expects that). This label has to be
global.
</para>

<para>
Lines 10-13 ask the system to write <varname>hbytes</varname> bytes
of the <varname>hello</varname> string to <varname>stdout</varname>.
</para>

<para>
Lines 15-16 ask the system to end the program with the return
value of <constant>0</constant>. The <function role="syscall">SYS_exit</function> syscall never
returns, so the code ends there.
</para>

<note>
<para>
If you have come to Unix from <acronym>MS DOS</acronym>
assembly language background, you may be used to writing directly
to the video hardware. You will never have to worry about
this in FreeBSD, or any other flavor of Unix. As far as
you are concerned, you are writing to a file known as
<filename>stdout</filename>. This can be the video screen, or
a <application>telnet</application> terminal, or an actual file,
or even the input of another program. Which one it is,
is for the system to figure out.
</para>
</note>

<sect2 id="x86-assemble-1"><title>Assembling the Code</title>

<para>
Type the code (except the line numbers) in an editor, and save
it in a file named <filename>hello.asm</filename>. You need
<application>nasm</application> to assemble it.
</para>

<sect3 id="x86-get-nasm"><title>Installing <application>nasm</application></title>

<para>
If you do not have <application>nasm</application>, type:
</para>

<screen>&prompt.user; <userinput>su</userinput>
Password:<userinput><replaceable>your root password</replaceable></userinput>
&prompt.root; <userinput>cd /usr/ports/devel/nasm</userinput>
&prompt.root; <userinput>make install</userinput>
&prompt.root; <userinput>exit</userinput>
&prompt.user;</screen>

<para>
You may type <userinput>make install clean</userinput> instead of just
<userinput>make install</userinput> if you do not want to keep
<application>nasm</application> source code.
</para>

<para>
Either way, FreeBSD will automatically download
<application>nasm</application> from the Internet,
compile it, and install it on your system.
</para>

<note>
<para>
If your system is not FreeBSD, you need to get
<application>nasm</application> from its
<ulink url="http://www.web-sites.co.uk/nasm/">home
page</ulink>. You can still use it to assemble FreeBSD code.
</para>
</note>

<para>
Now you can assemble, link, and run the code:
</para>

<screen>&prompt.user; <userinput>nasm -f elf hello.asm</userinput>
&prompt.user; <userinput>ld -s -o hello hello.o</userinput>
&prompt.user; <userinput>./hello</userinput>
Hello, World!
&prompt.user;</screen>

</sect3>

</sect2>

</sect1>

<sect1 id="x86-unix-filters">
<title>Writing Unix Filters</title>

<para>
A common type of Unix application is a filter&mdash;a program
that reads data from the <filename>stdin</filename>, processes it
somehow, then writes the result to <filename>stdout</filename>.
</para>

<para>
In this chapter, we shall develop a simple filter, and
learn how to read from <filename>stdin</filename> and write to
<filename>stdout</filename>. This filter will convert each byte
of its input into a hexadecimal number followed by a
blank space.
</para>

<programlisting>
%include	'system.inc'

section	.data
hex	db	'0123456789ABCDEF'
buffer	db	0, 0, ' '

section	.text
global	_start
_start:
	; read a byte from stdin
	push	dword 1
	push	dword buffer
	push	dword stdin
	sys.read
	add	esp, byte 12
	or	eax, eax
	je	.done

	; convert it to hex
	movzx	eax, byte [buffer]
	mov	edx, eax
	shr	dl, 4
	mov	dl, [hex+edx]
	mov	[buffer], dl
	and	al, 0Fh
	mov	al, [hex+eax]
	mov	[buffer+1], al

	; print it
	push	dword 3
	push	dword buffer
	push	dword stdout
	sys.write
	add	esp, byte 12
	jmp	short _start

.done:
	push	dword 0
	sys.exit
</programlisting>
<para>
In the data section we create an array called <varname>hex</varname>.
It contains the 16 hexadecimal digits in ascending order.
The array is followed by a buffer which we will use for
both input and output. The first two bytes of the buffer
are initially set to <constant>0</constant>. This is where we will write
the two hexadecimal digits (the first byte also is
where we will read the input). The third byte is a
space.
</para>

<para>
The code section consists of four parts: Reading the byte,
converting it to a hexadecimal number, writing the result,
and eventually exiting the program.
</para>

<para>
To read the byte, we ask the system to read one byte
from <filename>stdin</filename>, and store it in the first byte
of the <varname>buffer</varname>. The system returns the number
of bytes read in <varname role="register">EAX</varname>. This will be <constant>1</constant>
while data is coming, or <constant>0</constant>, when no more input
data is available. Therefore, we check the value of
<varname role="register">EAX</varname>. If it is <constant>0</constant>,
we jump to <varname>.done</varname>, otherwise we continue.
</para>

<note>
<para>
For simplicity sake, we are ignoring the possibility
of an error condition at this time.
</para>
</note>

<para>
The hexadecimal conversion reads the byte from the
<varname>buffer</varname> into <varname role="register">EAX</varname>, or actually just
<varname role="register">AL</varname>, while clearing the remaining bits of
<varname role="register">EAX</varname> to zeros. We also copy the byte to
<varname role="register">EDX</varname> because we need to convert the upper
four bits (nibble) separately from the lower
four bits. We store the result in the first two
bytes of the buffer.
</para>

<para>
Next, we ask the system to write the three bytes
of the buffer, i.e., the two hexadecimal digits and
the blank space, to <filename>stdout</filename>. We then
jump back to the beginning of the program and
process the next byte.
</para>

<para>
Once there is no more input left, we ask the system
to exit our program, returning a zero, which is
the traditional value meaning the program was
successful.
</para>

<para>
Go ahead, and save the code in a file named <filename>hex.asm</filename>,
then type the following (the <userinput>^D</userinput> means press the
control key and type <userinput>D</userinput> while holding the
control key down):
</para>

<screen>&prompt.user; <userinput>nasm -f elf hex.asm</userinput>
&prompt.user; <userinput>ld -s -o hex hex.o</userinput>
&prompt.user; <userinput>./hex</userinput>
<userinput>Hello, World!</userinput>
48 65 6C 6C 6F 2C 20 57 6F 72 6C 64 21 0A <userinput>Here I come!</userinput>
48 65 72 65 20 49 20 63 6F 6D 65 21 0A <userinput>^D</userinput> &prompt.user;</screen>

<note>
<para>
If you are migrating to Unix from <acronym>MS DOS</acronym>,
you may be wondering why each line ends with <constant>0A</constant>
instead of <constant>0D 0A</constant>.
This is because Unix does not use the cr/lf convention, but
a "new line" convention, which is <constant>0A</constant> in hexadecimal.
</para>
</note>

<para>
Can we improve this? Well, for one, it is a bit confusing because
once we have converted a line of text, our input no longer
starts at the begining of the line. We can modify it to print
a new line instead of a space after each <constant>0A</constant>:
</para>

<programlisting>
%include	'system.inc'

section	.data
hex	db	'0123456789ABCDEF'
buffer	db	0, 0, ' '

section	.text
global	_start
_start:
	mov	cl, ' '

.loop:
	; read a byte from stdin
	push	dword 1
	push	dword buffer
	push	dword stdin
	sys.read
	add	esp, byte 12
	or	eax, eax
	je	.done

	; convert it to hex
	movzx	eax, byte [buffer]
	mov	[buffer+2], cl
	cmp	al, 0Ah
	jne	.hex
	mov	[buffer+2], al

.hex:
	mov	edx, eax
	shr	dl, 4
	mov	dl, [hex+edx]
	mov	[buffer], dl
	and	al, 0Fh
	mov	al, [hex+eax]
	mov	[buffer+1], al

	; print it
	push	dword 3
	push	dword buffer
	push	dword stdout
	sys.write
	add	esp, byte 12
	jmp	short .loop

.done:
	push	dword 0
	sys.exit
</programlisting>
<para>
We have stored the space in the <varname role="register">CL</varname> register. We can
do this safely because, unlike Microsoft Windows, Unix system
calls do not modify the value of any register they do not use
to return a value in.
</para>

<para>
That means we only need to set <varname role="register">CL</varname> once. We have, therefore,
added a new label <varname>.loop</varname> and jump to it for the next byte
instead of jumping at <varname>_start</varname>. We have also added the
<varname>.hex</varname> label so we can either have a blank space or a
new line as the third byte of the <varname>buffer</varname>.
</para>

<para>
Once you have changed <filename>hex.asm</filename> to reflect
these changes, type:
</para>

<screen>&prompt.user; <userinput>nasm -f elf hex.asm</userinput>
&prompt.user; <userinput>ld -s -o hex hex.o</userinput>
&prompt.user; <userinput>./hex</userinput>
<userinput>Hello, World!</userinput>
48 65 6C 6C 6F 2C 20 57 6F 72 6C 64 21 0A
<userinput>Here I come!</userinput>
48 65 72 65 20 49 20 63 6F 6D 65 21 0A
<userinput>^D</userinput> &prompt.user;</screen>

<para>
That looks better. But this code is quite inefficient! We
are making a system call for every single byte twice (once
to read it, another time to write the output).
</para>

</sect1>

<sect1 id="x86-buffered-io">
<title>Buffered Input and Output</title>

<para>
We can improve the efficiency of our code by buffering our
input and output. We create an input buffer and read a whole
sequence of bytes at one time. Then we fetch them one by one
from the buffer.
</para>

<para>
We also create an output buffer. We store our output in it until
it is full. At that time we ask the kernel to write the contents
of the buffer to <filename>stdout</filename>.
</para>

<para>
The program ends when there is no more input. But we still need
to ask the kernel to write the contents of our output buffer
to <filename>stdout</filename> one last time, otherwise some of our output
would make it to the output buffer, but never be sent out.
Do not forget that, or you will be wondering why some of your
output is missing.
</para>

<programlisting>
%include	'system.inc'

%define	BUFSIZE	2048

section	.data
hex	db	'0123456789ABCDEF'

section .bss
ibuffer	resb	BUFSIZE
obuffer	resb	BUFSIZE

section	.text
global	_start
_start:
	sub	eax, eax
	sub	ebx, ebx
	sub	ecx, ecx
	mov	edi, obuffer

.loop:
	; read a byte from stdin
	call	getchar

	; convert it to hex
	mov	dl, al
	shr	al, 4
	mov	al, [hex+eax]
	call	putchar

	mov	al, dl
	and	al, 0Fh
	mov	al, [hex+eax]
	call	putchar

	mov	al, ' '
	cmp	dl, 0Ah
	jne	.put
	mov	al, dl

.put:
	call	putchar
	jmp	short .loop

align 4
getchar:
	or	ebx, ebx
	jne	.fetch

	call	read

.fetch:
	lodsb
	dec	ebx
	ret

read:
	push	dword BUFSIZE
	mov	esi, ibuffer
	push	esi
	push	dword stdin
	sys.read
	add	esp, byte 12
	mov	ebx, eax
	or	eax, eax
	je	.done
	sub	eax, eax
	ret

align 4
.done:
	call	write		; flush output buffer
	push	dword 0
	sys.exit

align 4
putchar:
	stosb
	inc	ecx
	cmp	ecx, BUFSIZE
	je	write
	ret

align 4
write:
	sub	edi, ecx	; start of buffer
	push	ecx
	push	edi
	push	dword stdout
	sys.write
	add	esp, byte 12
	sub	eax, eax
	sub	ecx, ecx	; buffer is empty now
	ret
</programlisting>
<para>
We now have a third section in the source code, named
<varname>.bss</varname>. This section is not included in our
executable file, and, therefore, cannot be initialized. We use
<function role="opcode">resb</function> instead of <function role="opcode">db</function>.
It simply reserves the requested size of uninitialized memory
for our use.
</para>

<para>
We take advantage of the fact that the system does not modify the
registers: We use registers for what, otherwise, would have to be
global variables stored in the <varname>.data</varname> section. This is
also why the Unix convention of passing parameters to system calls
on the stack is superior to the Microsoft convention of passing
them in the registers: We can keep the registers for our own use.
</para>

<para>
We use <varname role="register">EDI</varname> and <varname role="register">ESI</varname> as pointers to the next byte
to be read from or written to. We use <varname role="register">EBX</varname> and
<varname role="register">ECX</varname> to keep count of the number of bytes in the
two buffers, so we know when to dump the output to, or read more
input from, the system.
</para>

<para>
Let us see how it works now:
</para>

<screen>&prompt.user; <userinput>nasm -f elf hex.asm</userinput>
&prompt.user; <userinput>ld -s -o hex hex.o</userinput>
&prompt.user; <userinput>./hex</userinput>
<userinput>Hello, World!</userinput>
<userinput>Here I come!</userinput>
48 65 6C 6C 6F 2C 20 57 6F 72 6C 64 21 0A
48 65 72 65 20 49 20 63 6F 6D 65 21 0A
<userinput>^D</userinput> &prompt.user;</screen>

<para>
Not what you expected? The program did not print the output
until we pressed <userinput>^D</userinput>. That is easy to fix by
inserting three lines of code to write the output every time
we have converted a new line to <constant>0A</constant>. I have marked
the three lines with &gt; (do not copy the &gt; in your
<filename>hex.asm</filename>).
</para>

<programlisting>
%include	'system.inc'

%define	BUFSIZE	2048

section	.data
hex	db	'0123456789ABCDEF'

section .bss
ibuffer	resb	BUFSIZE
obuffer	resb	BUFSIZE

section	.text
global	_start
_start:
	sub	eax, eax
	sub	ebx, ebx
	sub	ecx, ecx
	mov	edi, obuffer

.loop:
	; read a byte from stdin
	call	getchar

	; convert it to hex
	mov	dl, al
	shr	al, 4
	mov	al, [hex+eax]
	call	putchar

	mov	al, dl
	and	al, 0Fh
	mov	al, [hex+eax]
	call	putchar

	mov	al, ' '
	cmp	dl, 0Ah
	jne	.put
	mov	al, dl

.put:
	call	putchar
>	cmp	al, 0Ah
>	jne	.loop
>	call	write
	jmp	short .loop

align 4
getchar:
	or	ebx, ebx
	jne	.fetch

	call	read

.fetch:
	lodsb
	dec	ebx
	ret

read:
	push	dword BUFSIZE
	mov	esi, ibuffer
	push	esi
	push	dword stdin
	sys.read
	add	esp, byte 12
	mov	ebx, eax
	or	eax, eax
	je	.done
	sub	eax, eax
	ret

align 4
.done:
	call	write		; flush output buffer
	push	dword 0
	sys.exit

align 4
putchar:
	stosb
	inc	ecx
	cmp	ecx, BUFSIZE
	je	write
	ret

align 4
write:
	sub	edi, ecx	; start of buffer
	push	ecx
	push	edi
	push	dword stdout
	sys.write
	add	esp, byte 12
	sub	eax, eax
	sub	ecx, ecx	; buffer is empty now
	ret
</programlisting>

<para>
Now, let us see how it works:
</para>

<screen>&prompt.user; <userinput>nasm -f elf hex.asm</userinput>
&prompt.user; <userinput>ld -s -o hex hex.o</userinput>
&prompt.user; <userinput>./hex</userinput>
<userinput>Hello, World!</userinput>
48 65 6C 6C 6F 2C 20 57 6F 72 6C 64 21 0A
<userinput>Here I come!</userinput>
48 65 72 65 20 49 20 63 6F 6D 65 21 0A
<userinput>^D</userinput> &prompt.user;</screen>

<para>
Not bad for a 644-byte executable, is it!
</para>

<note>
<para>
This approach to buffered input/output still
contains a hidden danger. I will discuss&mdash;and
fix&mdash;it later, when I talk about the
<link linkend="x86-buffered-dark-side">dark
side of buffering</link>.</para>
</note>

<sect2 id="x86-ungetc">
<title>How to Unread a Character</title>

<warning><para>
This may be a somewhat advanced topic, mostly of interest to
programmers familiar with the theory of compilers. If you wish,
you may <link linkend="x86-command-line">skip to the next
section</link>, and perhaps read this later.
</para>
</warning>
<para>
While our sample program does not require it, more sophisticated
filters often need to look ahead. In other words, they may need
to see what the next character is (or even several characters).
If the next character is of a certain value, it is part of the
token currently being processed. Otherwise, it is not.
</para>

<para>
For example, you may be parsing the input stream for a textual
string (e.g., when implementing a language compiler): If a
character is followed by another character, or perhaps a digit,
it is part of the token you are processing. If it is followed by
white space, or some other value, then it is not part of the
current token.
</para>

<para>
This presents an interesting problem: How to return the next
character back to the input stream, so it can be read again
later?
</para>

<para>
One possible solution is to store it in a character variable,
then set a flag. We can modify <function>getchar</function> to check the flag,
and if it is set, fetch the byte from that variable instead of the
input buffer, and reset the flag. But, of course, that slows us
down.
</para>

<para>
The C language has an <function>ungetc()</function> function, just for that
purpose. Is there a quick way to implement it in our code?
I would like you to scroll back up and take a look at the
<function>getchar</function> procedure and see if you can find a nice and
fast solution before reading the next paragraph. Then come back
here and see my own solution.
</para>

<para>
The key to returning a character back to the stream is in how
we are getting the characters to start with:
</para>

<para>
First we check if the buffer is empty by testing the value
of <varname role="register">EBX</varname>. If it is zero, we call the
<function>read</function> procedure.
</para>

<para>
If we do have a character available, we use <function role="opcode">lodsb</function>, then
decrease the value of <varname role="register">EBX</varname>. The <function role="opcode">lodsb</function>
instruction is effectively identical to:
</para>

<programlisting>
	mov	al, [esi]
	inc	esi
</programlisting>

<para>
The byte we have fetched remains in the buffer until the next
time <function>read</function> is called. We do not know when that happens,
but we do know it will not happen until the next call to
<function>getchar</function>. Hence, to "return" the last-read byte back
to the stream, all we have to do is decrease the value of
<varname role="register">ESI</varname> and increase the value of <varname role="register">EBX</varname>:
</para>

<programlisting>
ungetc:
	dec	esi
	inc	ebx
	ret
</programlisting>

<para>
But, be careful! We are perfectly safe doing this if our look-ahead
is at most one character at a time. If we are examining more than
one upcoming character and call <function>ungetc</function> several times
in a row, it will work most of the time, but not all the time
(and will be tough to debug). Why?
</para>

<para>
Because as long as <function>getchar</function> does not have to call
<function>read</function>, all of the pre-read bytes are still in the buffer,
and our <function>ungetc</function> works without a glitch. But the moment
<function>getchar</function> calls <function>read</function>,
the contents of the buffer change.
</para>

<para>
We can always rely on <function>ungetc</function> working properly on the last
character we have read with <function>getchar</function>, but not on anything
we have read before that.
</para>

<para>
If your program reads more than one byte ahead, you have at least
two choices:
</para>

<para>
If possible, modify the program so it only reads one byte ahead.
This is the simplest solution.
</para>

<para>
If that option is not available, first of all determine the maximum
number of characters your program needs to return to the input
stream at one time. Increase that number slightly, just to be
sure, preferably to a multiple of 16&mdash;so it aligns nicely.
Then modify the <varname>.bss</varname> section of your code, and create
a small "spare" buffer right before your input buffer,
something like this:
</para>

<programlisting>
section	.bss
	resb	16	; or whatever the value you came up with
ibuffer	resb	BUFSIZE
obuffer	resb	BUFSIZE
</programlisting>

<para>
You also need to modify your <function>ungetc</function> to pass the value
of the byte to unget in <varname role="register">AL</varname>:
</para>

<programlisting>
ungetc:
	dec	esi
	inc	ebx
	mov	[esi], al
	ret
</programlisting>

<para>
With this modification, you can call <function>ungetc</function>
up to 17 times in a row safely (the first call will still
be within the buffer, the remaining 16 may be either within
the buffer or within the "spare").
</para>

</sect2>

</sect1>

<sect1 id="x86-command-line"><title>Command Line Arguments</title>

<para>
Our <application>hex</application> program will be more useful if it can
read the names of an input and output file from its command
line, i.e., if it can process the command line arguments.
But... Where are they?
</para>

<para>
Before a Unix system starts a program, it <function role="opcode">push</function>es some
data on the stack, then jumps at the <varname>_start</varname>
label of the program. Yes, I said jumps, not calls. That means the
data can be accessed by reading <varname>[esp+offset]</varname>,
or by simply <function role="opcode">pop</function>ping it.
</para>

<para>
The value at the top of the stack contains the number of
command line arguments. It is traditionally called
<varname>argc</varname>, for "argument count."
</para>

<para>
Command line arguments follow next, all <varname>argc</varname> of them.
These are typically referred to as <varname>argv</varname>, for
"argument value(s)." That is, we get <varname>argv[0]</varname>,
<varname>argv[1]</varname>, <varname>...</varname>,
<varname>argv[argc-1]</varname>. These are not the actual
arguments, but pointers to arguments, i.e., memory addresses of
the actual arguments. The arguments themselves are
NUL-terminated character strings.
</para>

<para>
The <varname>argv</varname> list is followed by a NULL pointer,
which is simply a <constant>0</constant>. There is more, but this is
enough for our purposes right now.
</para>

<note>
<para>
If you have come from the <acronym>MS DOS</acronym> programming
environment, the main difference is that each argument is in
a separate string. The second difference is that there is no
practical limit on how many arguments there can be.
</para>
</note>

<para>
Armed with this knowledge, we are almost ready for the next
version of <filename>hex.asm</filename>. First, however, we need to
add a few lines to <filename>system.inc</filename>:
</para>

<para>
First, we need to add two new entries to our list of system
call numbers:
</para>

<programlisting>
%define	SYS_open	5
%define	SYS_close	6
</programlisting>

<para>
Then we add two new macros at the end of the file:
</para>

<programlisting>
%macro	sys.open	0
	system	SYS_open
%endmacro

%macro	sys.close	0
	system	SYS_close
%endmacro
</programlisting>

<para>
Here, then, is our modified source code:
</para>

<programlisting>
%include	'system.inc'

%define	BUFSIZE	2048

section	.data
fd.in	dd	stdin
fd.out	dd	stdout
hex	db	'0123456789ABCDEF'

section .bss
ibuffer	resb	BUFSIZE
obuffer	resb	BUFSIZE

section	.text
align 4
err:
	push	dword 1		; return failure
	sys.exit

align 4
global	_start
_start:
	add	esp, byte 8	; discard argc and argv[0]

	pop	ecx
	jecxz	.init		; no more arguments

	; ECX contains the path to input file
	push	dword 0		; O_RDONLY
	push	ecx
	sys.open
	jc	err		; open failed

	add	esp, byte 8
	mov	[fd.in], eax

	pop	ecx
	jecxz	.init		; no more arguments

	; ECX contains the path to output file
	push	dword 420	; file mode (644 octal)
	push	dword 0200h | 0400h | 01h
	; O_CREAT | O_TRUNC | O_WRONLY
	push	ecx
	sys.open
	jc	err

	add	esp, byte 12
	mov	[fd.out], eax

.init:
	sub	eax, eax
	sub	ebx, ebx
	sub	ecx, ecx
	mov	edi, obuffer

.loop:
	; read a byte from input file or stdin
	call	getchar

	; convert it to hex
	mov	dl, al
	shr	al, 4
	mov	al, [hex+eax]
	call	putchar

	mov	al, dl
	and	al, 0Fh
	mov	al, [hex+eax]
	call	putchar

	mov	al, ' '
	cmp	dl, 0Ah
	jne	.put
	mov	al, dl

.put:
	call	putchar
	cmp	al, dl
	jne	.loop
	call	write
	jmp	short .loop

align 4
getchar:
	or	ebx, ebx
	jne	.fetch

	call	read

.fetch:
	lodsb
	dec	ebx
	ret

read:
	push	dword BUFSIZE
	mov	esi, ibuffer
	push	esi
	push	dword [fd.in]
	sys.read
	add	esp, byte 12
	mov	ebx, eax
	or	eax, eax
	je	.done
	sub	eax, eax
	ret

align 4
.done:
	call	write		; flush output buffer

	; close files
	push	dword [fd.in]
	sys.close

	push	dword [fd.out]
	sys.close

	; return success
	push	dword 0
	sys.exit

align 4
putchar:
	stosb
	inc	ecx
	cmp	ecx, BUFSIZE
	je	write
	ret

align 4
write:
	sub	edi, ecx	; start of buffer
	push	ecx
	push	edi
	push	dword [fd.out]
	sys.write
	add	esp, byte 12
	sub	eax, eax
	sub	ecx, ecx	; buffer is empty now
	ret
</programlisting>

<para>
In our <varname>.data</varname> section we now have two new variables,
<varname>fd.in</varname> and <varname>fd.out</varname>. We store the input and
output file descriptors here.
</para>

<para>
In the <varname>.text</varname> section we have replaced the references
to <varname>stdin</varname> and <varname>stdout</varname> with
<varname>[fd.in]</varname> and <varname>[fd.out]</varname>.
</para>

<para>
The <varname>.text</varname> section now starts with a simple error
handler, which does nothing but exit the program with a return
value of <constant>1</constant>.
The error handler is before <varname>_start</varname> so we are
within a short distance from where the errors occur.
</para>

<para>
Naturally, the program execution still begins at <varname>_start</varname>.
First, we remove <varname>argc</varname> and <varname>argv[0]</varname> from the
stack: They are of no interest to us (in this program, that is).
</para>

<para>
We pop <varname>argv[1]</varname> to <varname role="register">ECX</varname>. This
register is particularly suited for pointers, as we can handle
NULL pointers with <function role="opcode">jecxz</function>. If <varname>argv[1]</varname>
is not NULL, we try to open the file named in the first
argument. Otherwise, we continue the program as before: Reading
from <varname>stdin</varname>, writing to <varname>stdout</varname>.
If we fail to open the input file (e.g., it does not exist),
we jump to the error handler and quit.
</para>

<para>
If all went well, we now check for the second argument. If
it is there, we open the output file. Otherwise, we send
the output to <varname>stdout</varname>. If we fail to open the output
file (e.g., it exists and we do not have the write permission),
we, again, jump to the error handler.
</para>

<para>
The rest of the code is the same as before, except we close
the input and output files before exiting, and, as mentioned,
we use <varname>[fd.in]</varname> and <varname>[fd.out]</varname>.
</para>

<para>
Our executable is now a whopping 768 bytes long.
</para>

<para>
Can we still improve it? Of course! Every program can be improved.
Here are a few ideas of what we could do:
</para>

<itemizedlist>
<listitem>
<para>
Have our error handler print a message to
<varname>stderr</varname>.
</para>
</listitem>

<listitem>
<para>
Add error handlers to the <function>read</function>
and <function>write</function> functions.
</para>
</listitem>

<listitem>
<para>
Close <varname>stdin</varname> when we open an input file,
<varname>stdout</varname> when we open an output file.
</para>
</listitem>

<listitem>
<para>
Add command line switches, such as <parameter>-i</parameter>
and <parameter>-o</parameter>, so we can list the input and
output files in any order, or perhaps read from
<varname>stdin</varname> and write to a file.
</para>
</listitem>

<listitem>
<para>
Print a usage message if command line arguments are incorrect.
</para>
</listitem>

</itemizedlist>
<para>
I shall leave these enhancements as an exercise to the reader:
You already know everything you need to know to implement them.
</para>

</sect1>

<sect1 id="x86-environment">
<title>Unix Environment</title>

<para>
An important Unix concept is the environment, which is defined by
<emphasis>environment variables</emphasis>. Some are set by the system, others
by you, yet others by the <application>shell</application>, or any program
that loads another program.
</para>

<sect2 id="x86-find-environment">
<title>How to Find Environment Variables</title>

<para>
I said earlier that when a program starts executing, the stack
contains <varname>argc</varname> followed by the NULL-terminated
<varname>argv</varname> array, followed by something else. The
"something else" is the <emphasis>environment</emphasis>, or,
to be more precise, a NULL-terminated array of pointers to
<emphasis>environment variables</emphasis>. This is often referred
to as <varname>env</varname>.
</para>

<para>
The structure of <varname>env</varname> is the same as that of
<varname>argv</varname>, a list of memory addresses followed by a
NULL (<constant>0</constant>). In this case, there is no
<varname>"envc"</varname>&mdash;we figure out where the array ends
by searching for the final NULL.
</para>

<para>
The variables usually come in the <varname>name=value</varname>
format, but sometimes the <varname>=value</varname> part
may be missing. We need to account for that possibility.
</para>

</sect2>

<sect2 id="x86-webvar">
<title>webvars</title>

<para>
I could just show you some code that prints the environment
the same way the Unix <application>env</application> command does. But
I thought it would be more interesting to write a simple
assembly language CGI utility.
</para>

<sect3 id="x86-cgi">
<title>CGI: A Quick Overview</title>

<para>
I have a
<ulink url="http://www.whizkidtech.net/cgi-bin/tutorial">detailed
<acronym>CGI</acronym> tutorial</ulink> on my web site,
but here is a very quick overview of <acronym>CGI</acronym>:
</para>

<itemizedlist>
<listitem>
<para>
The web server communicates with the <acronym>CGI</acronym>
program by setting <emphasis>environment variables</emphasis>.
</para>
</listitem>

<listitem>
<para>
The <acronym>CGI</acronym> program
sends its output to <filename>stdout</filename>.
The web server reads it from there.
</para>
</listitem>

<listitem>
<para>
It must start with an <acronym>HTTP</acronym>
header followed by two blank lines.
</para>
</listitem>

<listitem>
<para>
It then prints the <acronym>HTML</acronym>
code, or whatever other type of data it is producing.
</para>
</listitem>

</itemizedlist>
<note>
<para>
While certain <emphasis>environment variables</emphasis> use
standard names, others vary, depending on the web server. That
makes <application>webvars</application>
quite a useful diagnostic tool.
</para>
</note>

</sect3>

<sect3 id="x86-webvars-the-code">
<title>The Code</title>

<para>
Our <application>webvars</application> program, then, must send out
the <acronym>HTTP</acronym> header followed by some
<acronym>HTML</acronym> mark-up. It then must read
the <emphasis>environment variables</emphasis> one by one
and send them out as part of the
<acronym>HTML</acronym> page.
</para>

<para>
The code follows. I placed comments and explanations
right inside the code:
</para>

<programlisting>
;;;;;;; webvars.asm ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Copyright (c) 2000 G. Adam Stanislav
; All rights reserved.
;
; Redistribution and use in source and binary forms, with or without
; modification, are permitted provided that the following conditions
; are met:
; 1. Redistributions of source code must retain the above copyright
;    notice, this list of conditions and the following disclaimer.
; 2. Redistributions in binary form must reproduce the above copyright
;    notice, this list of conditions and the following disclaimer in the
;    documentation and/or other materials provided with the distribution.
;
; THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
; ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
; IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
; ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
; FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
; DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
; OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
; HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
; LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
; OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
; SUCH DAMAGE.
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Version 1.0
;
; Started:	 8-Dec-2000
; Updated:	 8-Dec-2000
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
%include	'system.inc'

section	.data
http	db	'Content-type: text/html', 0Ah, 0Ah
	db	'&lt;?xml version="1.0" encoding="UTF-8"?&gt;', 0Ah
	db	'&lt;!DOCTYPE html PUBLIC "-//W3C/DTD XHTML Strict//EN" '
	db	'"DTD/xhtml1-strict.dtd"&gt;', 0Ah
	db	'&lt;html xmlns="http://www.w3.org/1999/xhtml" '
	db	'xml.lang="en" lang="en"&gt;', 0Ah
	db	'&lt;head&gt;', 0Ah
	db	'&lt;title&gt;Web Environment&lt;/title&gt;', 0Ah
	db	'&lt;meta name="author" content="G. Adam Stanislav" /&gt;', 0Ah
	db	'&lt;/head&gt;', 0Ah, 0Ah
	db	'&lt;body bgcolor="#ffffff" text="#000000" link="#0000ff" '
	db	'vlink="#840084" alink="#0000ff"&gt;', 0Ah
	db	'&lt;div class="webvars"&gt;', 0Ah
	db	'&lt;h1&gt;Web Environment&lt;/h1&gt;', 0Ah
	db	'&lt;p&gt;The following &lt;b&gt;environment variables&lt;/b&gt; are defined '
	db	'on this web server:&lt;/p&gt;', 0Ah, 0Ah
	db	'&lt;table align="center" width="80" border="0" cellpadding="10" '
	db	'cellspacing="0" class="webvars"&gt;', 0Ah
httplen	equ	$-http
left	db	'&lt;tr&gt;', 0Ah
	db	'&lt;td class="name"&gt;&lt;tt&gt;'
leftlen	equ	$-left
middle	db	'&lt;/tt&gt;&lt;/td&gt;', 0Ah
	db	'&lt;td class="value"&gt;&lt;tt&gt;&lt;b&gt;'
midlen	equ	$-middle
undef	db	'&lt;i&gt;(undefined)&lt;/i&gt;'
undeflen	equ	$-undef
right	db	'&lt;/b&gt;&lt;/tt&gt;&lt;/td&gt;', 0Ah
	db	'&lt;/tr&gt;', 0Ah
rightlen	equ	$-right
wrap	db	'&lt;/table&gt;', 0Ah
	db	'&lt;/div&gt;', 0Ah
	db	'&lt;/body&gt;', 0Ah
	db	'&lt;/html&gt;', 0Ah, 0Ah
wraplen	equ	$-wrap

section	.text
global	_start
_start:
	; First, send out all the http and xhtml stuff that is
	; needed before we start showing the environment
	push	dword httplen
	push	dword http
	push	dword stdout
	sys.write

	; Now find how far on the stack the environment pointers
	; are. We have 12 bytes we have pushed before "argc"
	mov	eax, [esp+12]

	; We need to remove the following from the stack:
	;
	;	The 12 bytes we pushed for sys.write
	;	The  4 bytes of argc
	;	The EAX*4 bytes of argv
	;	The  4 bytes of the NULL after argv
	;
	; Total:
	;	20 + eax * 4
	;
	; Because stack grows down, we need to ADD that many bytes
	; to ESP.
	lea	esp, [esp+20+eax*4]
	cld		; This should already be the case, but let's be sure.

	; Loop through the environment, printing it out
.loop:
	pop	edi
	or	edi, edi	; Done yet?
	je	near .wrap

	; Print the left part of HTML
	push	dword leftlen
	push	dword left
	push	dword stdout
	sys.write

	; It may be tempting to search for the '=' in the env string next.
	; But it is possible there is no '=', so we search for the
	; terminating NUL first.
	mov	esi, edi	; Save start of string
	sub	ecx, ecx
	not	ecx		; ECX = FFFFFFFF
	sub	eax, eax
repne	scasb
	not	ecx		; ECX = string length + 1
	mov	ebx, ecx	; Save it in EBX

	; Now is the time to find '='
	mov	edi, esi	; Start of string
	mov	al, '='
repne	scasb
	not	ecx
	add	ecx, ebx	; Length of name

	push	ecx
	push	esi
	push	dword stdout
	sys.write

	; Print the middle part of HTML table code
	push	dword midlen
	push	dword middle
	push	dword stdout
	sys.write

	; Find the length of the value
	not	ecx
	lea	ebx, [ebx+ecx-1]

	; Print "undefined" if 0
	or	ebx, ebx
	jne	.value

	mov	ebx, undeflen
	mov	edi, undef

.value:
	push	ebx
	push	edi
	push	dword stdout
	sys.write

	; Print the right part of the table row
	push	dword rightlen
	push	dword right
	push	dword stdout
	sys.write

	; Get rid of the 60 bytes we have pushed
	add	esp, byte 60

	; Get the next variable
	jmp	.loop

.wrap:
	; Print the rest of HTML
	push	dword wraplen
	push	dword wrap
	push	dword stdout
	sys.write

	; Return success
	push	dword 0
	sys.exit
</programlisting>

<para>
This code produces a 1,396-byte executable. Most of it is data,
i.e., the <acronym>HTML</acronym> mark-up we need to send out.
</para>

<para>
Assemble and link it as usual:
</para>

<screen>&prompt.user; <userinput>nasm -f elf webvars.asm</userinput>
&prompt.user; <userinput>ld -s -o webvars webvars.o</userinput></screen>

<para>
To use it, you need to upload <filename>webvars</filename> to your
web server. Depending on how your web server is set up, you
may have to store it in a special <filename>cgi-bin</filename> directory,
or perhaps rename it with a <filename>.cgi</filename> extension.
</para>

<para>
Then you need to use your browser to view its output.
To see its output on my web server, please go to
<ulink url="http://www.int80h.org/webvars/"><filename>http://www.int80h.org/webvars/</filename></ulink>.
If curious about the additional environment variables
present in a password protected web directory, go to
<ulink url="http://www.int80h.org/private/"><filename>http://www.int80h.org/private/</filename></ulink>,
using the name <userinput>asm</userinput> and password
<userinput>programmer</userinput>.
</para>

</sect3>

</sect2>

</sect1>

<sect1 id="x86-files">
<title>Working with Files</title>

<para>
We have already done some basic file work: We know how
to open and close them, how to read and write them using
buffers. But Unix offers much more functionality when it
comes to files. We will examine some of it in this section,
and end up with a nice file conversion utility.
</para>

<para>
Indeed, let us start at the end, that is, with the file
conversion utility. It always makes programming easier
when we know from the start what the end product is
supposed to do.
</para>

<para>
One of the first programs I wrote for Unix was
<ulink url="ftp://ftp.int80h.org/unix/tuc/"><application>tuc</application></ulink>,
a text-to-Unix file converter. It converts a text
file from other operating systems to a Unix text file.
In other words, it changes from different kind of line endings
to the newline convention of Unix. It saves the output
in a different file. Optionally, it converts a Unix text
file to a <acronym>DOS</acronym> text file.
</para>

<para>
I have used <application>tuc</application> extensively, but always
only to convert from some other <acronym>OS</acronym>
to Unix, never the other way. I have always wished
it would just overwrite the file instead of me having
to send the output to a different file. Most of the time,
I end up using it like this:
</para>

<screen>&prompt.user; <userinput>tuc <replaceable>myfile tempfile</replaceable></userinput>
&prompt.user; <userinput>mv <replaceable>tempfile myfile</replaceable></userinput></screen>

<para>
It would be nice to have a <application>ftuc</application>,
i.e., <emphasis>fast tuc</emphasis>, and use it like this:
</para>

<screen>&prompt.user; <userinput>ftuc <replaceable>myfile</replaceable></userinput></screen>

<para>
In this chapter, then, we will write
<application>ftuc</application> in assembly language
(the original <application>tuc</application>
is in C), and study various
file-oriented kernel services in the process.
</para>

<para>
At first sight, such a file conversion is very
simple: All you have to do is strip the carriage
returns, right?
</para>

<para>
If you answered yes, think again: That approach will
work most of the time (at least with <acronym>MS
DOS</acronym> text files), but will fail occasionally.
</para>

<para>
The problem is that not all non-Unix text files end their
line with the carriage return / line feed sequence. Some
use carriage returns without line feeds. Others combine several
blank lines into a single carriage return followed by several
line feeds. And so on.
</para>

<para>
A text file converter, then, must be able to handle
any possible line endings:
</para>

<itemizedlist>
<listitem>
<para>
carriage return / line feed
</para>
</listitem>

<listitem>
<para>
carriage return
</para>
</listitem>

<listitem>
<para>
line feed / carriage return
</para>
</listitem>

<listitem>
<para>
line feed
</para>
</listitem>

</itemizedlist>
<para>
It should also handle files that use some kind of a
combination of the above (e.g., carriage return followed
by several line feeds).
</para>

<sect2 id="x86-finite-state-machine">
<title>Finite State Machine</title>

<para>
The problem is easily solved by the use of a technique
called <emphasis>finite state machine</emphasis>, originally developed
by the designers of digital electronic circuits. A
<emphasis>finite state machine</emphasis> is a digital circuit
whose output is dependent not only on its input but on
its previous input, i.e., on its state. The microprocessor
is an example of a <emphasis>finite state machine</emphasis>: Our
assembly language code is assembled to machine language in which
some assembly language code produces a single byte
of machine language, while others produce several bytes.
As the microprocessor fetches the bytes from the memory
one by one, some of them simply change its state rather than
produce some output. When all the bytes of the op code are
fetched, the microprocessor produces some output, or changes
the value of a register, etc.
</para>

<para>
Because of that, all software is essentially a sequence of state
instructions for the microprocessor. Nevertheless, the concept
of <emphasis>finite state machine</emphasis> is useful in software design as well.
</para>

<para>
Our text file converter can be designed as a <emphasis>finite state machine</emphasis> with three
possible states. We could call them states 0-2,
but it will make our life easier if we give them symbolic names:
</para>

<itemizedlist>
<listitem>
<para>
<symbol>ordinary
</symbol></para>
</listitem>

<listitem>
<para>
<symbol>cr
</symbol></para>
</listitem>

<listitem>
<para>
<symbol>lf
</symbol></para>
</listitem>

</itemizedlist>
<para>
Our program will start in the <symbol>ordinary</symbol>
state. During this state, the program action depends on
its input as follows:
</para>

<itemizedlist>
<listitem>
<para>
If the input is anything other than a carriage return
or line feed, the input is simply passed on to the output. The
state remains unchanged.
</para>
</listitem>

<listitem>
<para>
If the input is a carriage return, the state is changed
to <symbol>cr</symbol>. The input is then discarded, i.e.,
no output is made.
</para>
</listitem>

<listitem>
<para>
If the input is a line feed, the state is changed to
<symbol>lf</symbol>. The input is then discarded.
</para>
</listitem>

</itemizedlist>
<para>
Whenever we are in the <symbol>cr</symbol> state, it is
because the last input was a carriage return, which was
unprocessed. What our software does in this state again
depends on the current input:
</para>

<itemizedlist>
<listitem>
<para>
If the input is anything other than a carriage return
or line feed, output a line feed, then output the input, then
change the state to <symbol>ordinary</symbol>.
</para>
</listitem>

<listitem>
<para>
If the input is a carriage return, we have received
two (or more) carriage returns in a row. We discard the
input, we output a line feed, and leave the state unchanged.
</para>
</listitem>

<listitem>
<para>
If the input is a line feed, we output the line feed
and change the state to <symbol>ordinary</symbol>. Note that
this is not the same as the first case above &ndash; if we tried
to combine them, we would be outputting two line feeds
instead of one.
</para>
</listitem>

</itemizedlist>
<para>
Finally, we are in the <symbol>lf</symbol> state after
we have received a line feed that was not preceded by a
carriage return. This will happen when our file already is
in Unix format, or whenever several lines in a row are
expressed by a single carriage return followed by several
line feeds, or when line ends with a line feed /
carriage return sequence. Here is how we need to handle
our input in this state:
</para>

<itemizedlist>
<listitem>
<para>
If the input is anything other than a carriage return or
line feed, we output a line feed, then output the input, then
change the state to <symbol>ordinary</symbol>. This is exactly
the same action as in the <symbol>cr</symbol> state upon
receiving the same kind of input.
</para>
</listitem>

<listitem>
<para>
If the input is a carriage return, we discard the input,
we output a line feed, then change the state to <symbol>ordinary</symbol>.
</para>
</listitem>

<listitem>
<para>
If the input is a line feed, we output the line feed,
and leave the state unchanged.
</para>
</listitem>

</itemizedlist>
<sect3 id="x86-final-state">
<title>The Final State</title>

<para>
The above <emphasis>finite state machine</emphasis> works for the entire file, but leaves
the possibility that the final line end will be ignored. That will
happen whenever the file ends with a single carriage return or
a single line feed. I did not think of it when I wrote
<application>tuc</application>, just to discover that
occasionally it strips the last line ending.
</para>

<para>
This problem is easily fixed by checking the state after the
entire file was processed. If the state is not
<symbol>ordinary</symbol>, we simply
need to output one last line feed.
</para>

<note>
<para>
Now that we have expressed our algorithm as a <emphasis>finite state machine</emphasis>,
we could easily design a dedicated digital electronic
circuit (a "chip") to do the conversion for us. Of course,
doing so would be considerably more expensive than writing
an assembly language program.
</para>
</note>

</sect3>

<sect3 id="x86-tuc-counter">
<title>The Output Counter</title>

<para>
Because our file conversion program may be combining two
characters into one, we need to use an output counter. We
initialize it to <constant>0</constant>, and increase it
every time we send a character to the output. At the end of
the program, the counter will tell us what size we need
to set the file to.
</para>

</sect3>

</sect2>

<sect2 id="x86-software-fsm">
<title>Implementing FSM in Software</title>

<para>
The hardest part of working with a <emphasis>finite state machine</emphasis>
is analyzing the problem and expressing it as a
<emphasis>finite state machine</emphasis>. That accomplished,
the software almost writes itself.
</para>

<para>
In a high-level language, such as C, there are several main
approaches. One is to use a <function role="statement">switch</function> statement
which chooses what function should be run. For example,
</para>

<programlisting>
	switch (state) {
	default:
	case REGULAR:
		regular(inputchar);
		break;
	case CR:
		cr(inputchar);
		break;
	case LF:
		lf(inputchar);
		break;
	}
</programlisting>

<para>
Another approach is by using an array of function pointers,
something like this:
</para>

<programlisting>
	(output[state])(inputchar);
</programlisting>

<para>
Yet another is to have <varname>state</varname> be a
function pointer, set to point at the appropriate function:
</para>

<programlisting>
	(*state)(inputchar);
</programlisting>
<para>
This is the approach we will use in our program because it is very easy to do in assembly language, and very fast, too. We will simply keep the address of the right procedure in <varname role="register">EBX</varname>, and then just issue:</para>

<programlisting>
	call	ebx
</programlisting>

<para>
This is possibly faster than hardcoding the address in the code
because the microprocessor does not have to fetch the address from
the memory&mdash;it is already stored in one of its registers. I said
<emphasis>possibly</emphasis> because with the caching modern
microprocessors do, either way may be equally fast.
</para>

</sect2>

<sect2 id="memory-mapped-files">
<title>Memory Mapped Files</title>

<para>
Because our program works on a single file, we cannot use the
approach that worked for us before, i.e., to read from an input
file and to write to an output file.
</para>

<para>
Unix allows us to map a file, or a section of a file,
into memory. To do that, we first need to open the file with the
appropriate read/write flags. Then we use the <function role="syscall">mmap</function>
system call to map it into the memory. One nice thing about
<function role="syscall">mmap</function> is that it automatically works with
virtual memory: We can map more of the file into the memory than
we have physical memory available, yet still access it through
regular memory op codes, such as <function role="opcode">mov</function>,
<function role="opcode">lods</function>, and <function role="opcode">stos</function>.
Whatever changes we make to the memory image of the file will be
written to the file by the system. We do not even have to keep
the file open: As long as it stays mapped, we can
read from it and write to it.
</para>

<para>
The 32-bit Intel microprocessors can access up to four
gigabytes of memory &ndash; physical or virtual. The FreeBSD system
allows us to use up to a half of it for file mapping.
</para>

<para>
For simplicity sake, in this tutorial we will only convert files
that can be mapped into the memory in their entirety. There are
probably not too many text files that exceed two gigabytes in size.
If our program encounters one, it will simply display a message
suggesting we use the original
<application>tuc</application> instead.
</para>

<para>
If you examine your copy of <filename>syscalls.master</filename>,
you will find two separate syscalls named <function role="syscall">mmap</function>.
This is because of evolution of Unix: There was the traditional
<acronym>BSD</acronym> <function role="syscall">mmap</function>,
syscall 71. That one was superceded by the <acronym>POSIX</acronym> <function role="syscall">mmap</function>,
syscall 197. The FreeBSD system supports both because
older programs were written by using the original <acronym>BSD</acronym>
version. But new software uses the <acronym>POSIX</acronym> version,
which is what we will use.
</para>

<para>
The <filename>syscalls.master</filename> file lists
the <acronym>POSIX</acronym> version like this:
</para>

<programlisting>
197	STD	BSD	{ caddr_t mmap(caddr_t addr, size_t len, int prot, \
			    int flags, int fd, long pad, off_t pos); }
</programlisting>

<para>
This differs slightly from what
<citerefentry><refentrytitle>mmap</refentrytitle><manvolnum>2</manvolnum></citerefentry>
says. That is because
<citerefentry><refentrytitle>mmap</refentrytitle><manvolnum>2</manvolnum></citerefentry>
describes the C version.
</para>

<para>
The difference is in the <varname>long pad</varname> argument, which is not present in the C version. However, the FreeBSD syscalls add a 32-bit pad after <function role="opcode">push</function>ing a 64-bit argument. In this case, <varname>off_t</varname> is a 64-bit value.</para>

<para>
When we are finished working with a memory-mapped file,
we unmap it with the <function role="syscall">munmap</function> syscall:
</para>

<tip>
<para>
For an in-depth treatment of <function role="syscall">mmap</function>, see
W. Richard Stevens'
<ulink url="http://www.int80h.org/cgi-bin/isbn?isbn=0130810819">Unix
Network Programming, Volume 2, Chapter 12</ulink>.
</para>
</tip>

</sect2>

<sect2 id="x86-file-size">
<title>Determining File Size</title>

<para>
Because we need to tell <function role="syscall">mmap</function> how many bytes
of the file to map into the memory, and because we want to map
the entire file, we need to determine the size of the file.
</para>

<para>
We can use the <function role="syscall">fstat</function> syscall to get all
the information about an open file that the system can give us.
That includes the file size.
</para>

<para>
Again, <filename>syscalls.master</filename> lists two versions
of <function role="syscall">fstat</function>, a traditional one
(syscall 62), and a <acronym>POSIX</acronym> one
(syscall 189). Naturally, we will use the
<acronym>POSIX</acronym> version:
</para>

<programlisting>
189	STD	POSIX	{ int fstat(int fd, struct stat *sb); }
</programlisting>

<para>
This is a very straightforward call: We pass to it the address
of a <structname>stat</structname> structure and the descriptor
of an open file. It will fill out the contents of the
<structname>stat</structname> structure.
</para>

<para>
I do, however, have to say that I tried to declare the
<structname>stat</structname> structure in the
<varname>.bss</varname> section, and
<function role="syscall">fstat</function> did not like it: It set the carry
flag indicating an error. After I changed the code to allocate
the structure on the stack, everything was working fine.
</para>

</sect2>

<sect2 id="x86-ftruncate">
<title>Changing the File Size</title>

<para>
Because our program may combine carriage return / line feed
sequences into straight line feeds, our output may be smaller
than our input. However, since we are placing our output into
the same file we read the input from, we may have to change the
size of the file.
</para>

<para>
The <function role="syscall">ftruncate</function> system call allows us to do
just that. Despite its somewhat misleading name, the
<function role="syscall">ftruncate</function> system call can be used to both
truncate the file (make it smaller) and to grow it.
</para>

<para>
And yes, we will find two versions of <function role="syscall">ftruncate</function>
in <filename>syscalls.master</filename>, an older one
(130), and a newer one (201). We will use
the newer one:
</para>

<programlisting>
201	STD	BSD	{ int ftruncate(int fd, int pad, off_t length); }
</programlisting>

<para>
Please note that this one contains a <varname>int pad</varname> again.
</para>

</sect2>

<sect2 id="x86-ftuc">
<title>ftuc</title>

<para>
We now know everything we need to write <application>ftuc</application>.
We start by adding some new lines in <filename>system.inc</filename>.
First, we define some constants and structures, somewhere at
or near the beginning of the file:
</para>

<programlisting>
;;;;;;; open flags
%define	O_RDONLY	0
%define	O_WRONLY	1
%define	O_RDWR	2

;;;;;;; mmap flags
%define	PROT_NONE	0
%define	PROT_READ	1
%define	PROT_WRITE	2
%define	PROT_EXEC	4
;;
%define	MAP_SHARED	0001h
%define	MAP_PRIVATE	0002h

;;;;;;; stat structure
struc	stat
st_dev		resd	1	; = 0
st_ino		resd	1	; = 4
st_mode		resw	1	; = 8, size is 16 bits
st_nlink	resw	1	; = 10, ditto
st_uid		resd	1	; = 12
st_gid		resd	1	; = 16
st_rdev		resd	1	; = 20
st_atime	resd	1	; = 24
st_atimensec	resd	1	; = 28
st_mtime	resd	1	; = 32
st_mtimensec	resd	1	; = 36
st_ctime	resd	1	; = 40
st_ctimensec	resd	1	; = 44
st_size		resd	2	; = 48, size is 64 bits
st_blocks	resd	2	; = 56, ditto
st_blksize	resd	1	; = 64
st_flags	resd	1	; = 68
st_gen		resd	1	; = 72
st_lspare	resd	1	; = 76
st_qspare	resd	4	; = 80
endstruc
</programlisting>

<para>
We define the new syscalls:
</para>

<programlisting>
%define	SYS_mmap	197
%define	SYS_munmap	73
%define	SYS_fstat	189
%define	SYS_ftruncate	201
</programlisting>

<para>
We add the macros for their use:
</para>

<programlisting>
%macro	sys.mmap	0
	system	SYS_mmap
%endmacro

%macro	sys.munmap	0
	system	SYS_munmap
%endmacro

%macro	sys.ftruncate	0
	system	SYS_ftruncate
%endmacro

%macro	sys.fstat	0
	system	SYS_fstat
%endmacro
</programlisting>

<para>
And here is our code:
</para>

<programlisting>
;;;;;;; Fast Text-to-Unix Conversion (ftuc.asm) ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;
;; Started:	21-Dec-2000
;; Updated:	22-Dec-2000
;;
;; Copyright 2000 G. Adam Stanislav.
;; All rights reserved.
;;
;;;;;;; v.1 ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
%include	'system.inc'

section	.data
	db	'Copyright 2000 G. Adam Stanislav.', 0Ah
	db	'All rights reserved.', 0Ah
usg	db	'Usage: ftuc filename', 0Ah
usglen	equ	$-usg
co	db	"ftuc: Can't open file.", 0Ah
colen	equ	$-co
fae	db	'ftuc: File access error.', 0Ah
faelen	equ	$-fae
ftl	db	'ftuc: File too long, use regular tuc instead.', 0Ah
ftllen	equ	$-ftl
mae	db	'ftuc: Memory allocation error.', 0Ah
maelen	equ	$-mae

section	.text

align 4
memerr:
	push	dword maelen
	push	dword mae
	jmp	short error

align 4
toolong:
	push	dword ftllen
	push	dword ftl
	jmp	short error

align 4
facerr:
	push	dword faelen
	push	dword fae
	jmp	short error

align 4
cantopen:
	push	dword colen
	push	dword co
	jmp	short error

align 4
usage:
	push	dword usglen
	push	dword usg

error:
	push	dword stderr
	sys.write

	push	dword 1
	sys.exit

align 4
global	_start
_start:
	pop	eax		; argc
	pop	eax		; program name
	pop	ecx		; file to convert
	jecxz	usage

	pop	eax
	or	eax, eax	; Too many arguments?
	jne	usage

	; Open the file
	push	dword O_RDWR
	push	ecx
	sys.open
	jc	cantopen

	mov	ebp, eax	; Save fd

	sub	esp, byte stat_size
	mov	ebx, esp

	; Find file size
	push	ebx
	push	ebp		; fd
	sys.fstat
	jc	facerr

	mov	edx, [ebx + st_size + 4]

	; File is too long if EDX != 0 ...
	or	edx, edx
	jne	near toolong
	mov	ecx, [ebx + st_size]
	; ... or if it is above 2 GB
	or	ecx, ecx
	js	near toolong

	; Do nothing if the file is 0 bytes in size
	jecxz	.quit

	; Map the entire file in memory
	push	edx
	push	edx		; starting at offset 0
	push	edx		; pad
	push	ebp		; fd
	push	dword MAP_SHARED
	push	dword PROT_READ | PROT_WRITE
	push	ecx		; entire file size
	push	edx		; let system decide on the address
	sys.mmap
	jc	near memerr

	mov	edi, eax
	mov	esi, eax
	push	ecx		; for SYS_munmap
	push	edi

	; Use EBX for state machine
	mov	ebx, ordinary
	mov	ah, 0Ah
	cld

.loop:
	lodsb
	call	ebx
	loop	.loop

	cmp	ebx, ordinary
	je	.filesize

	; Output final lf
	mov	al, ah
	stosb
	inc	edx

.filesize:
	; truncate file to new size
	push	dword 0		; high dword
	push	edx		; low dword
	push	eax		; pad
	push	ebp
	sys.ftruncate

	; close it (ebp still pushed)
	sys.close

	add	esp, byte 16
	sys.munmap

.quit:
	push	dword 0
	sys.exit

align 4
ordinary:
	cmp	al, 0Dh
	je	.cr

	cmp	al, ah
	je	.lf

	stosb
	inc	edx
	ret

align 4
.cr:
	mov	ebx, cr
	ret

align 4
.lf:
	mov	ebx, lf
	ret

align 4
cr:
	cmp	al, 0Dh
	je	.cr

	cmp	al, ah
	je	.lf

	xchg	al, ah
	stosb
	inc	edx

	xchg	al, ah
	; fall through

.lf:
	stosb
	inc	edx
	mov	ebx, ordinary
	ret

align 4
.cr:
	mov	al, ah
	stosb
	inc	edx
	ret

align 4
lf:
	cmp	al, ah
	je	.lf

	cmp	al, 0Dh
	je	.cr

	xchg	al, ah
	stosb
	inc	edx

	xchg	al, ah
	stosb
	inc	edx
	mov	ebx, ordinary
	ret

align 4
.cr:
	mov	ebx, ordinary
	mov	al, ah
	; fall through

.lf:
	stosb
	inc	edx
	ret
</programlisting>

<warning><para>
Do not use this program on files stored on a disk formated
by <acronym>MS DOS</acronym> or Windows. There seems to be a
subtle bug in the FreeBSD code when using <function role="syscall">mmap</function>
on these drives mounted under FreeBSD: If the file is over
a certain size, <function role="syscall">mmap</function> will just fill the memory
with zeros, and then copy them to the file overwriting
its contents.
</para>
</warning>
</sect2>

</sect1>

<sect1 id="x86-one-pointed-mind">
<title>One-Pointed Mind</title>

<para>
As a student of Zen, I like the idea of a one-pointed mind:
Do one thing at a time, and do it well.
</para>

<para>
This, indeed, is very much how Unix works as well. While
a typical Windows application is attempting to do everything
imaginable (and is, therefore, riddled with bugs), a
typical Unix program does only one thing, and it does it
well.
</para>

<para>
The typical Unix user then essentially assembles his own
applications by writing a shell script which combines the
various existing programs by piping the output of one
program to the input of another.
</para>

<para>
When writing your own Unix software, it is generally a
good idea to see what parts of the problem you need to
solve can be handled by existing programs, and only
write your own programs for that part of the problem
that you do not have an existing solution for.
</para>

<sect2 id="x86-csv"><title>CSV</title>

<para>
I will illustrate this principle with a specific real-life
example I was faced with recently:
</para>

<para>
I needed to extract the 11th field of each record from a
database I downloaded from a web site. The database was a
<acronym>CSV</acronym> file, i.e., a list of
<emphasis>comma-separated values</emphasis>. That is quite
a standard format for sharing data among people who may be
using different database software.
</para>

<para>
The first line of the file contains the list of various fields
separated by commas. The rest of the file contains the data
listed line by line, with values separated by commas.
</para>

<para>
I tried <application>awk</application>, using the comma as a separator.
But because several lines contained a quoted comma,
<application>awk</application> was extracting the wrong field
from those lines.
</para>

<para>
Therefore, I needed to write my own software to extract the 11th
field from the <acronym>CSV</acronym> file. However, going with the Unix
spirit, I only needed to write a simple filter that would do the
following:
</para>

<itemizedlist>
<listitem>
<para>
Remove the first line from the file;
</para>
</listitem>

<listitem>
<para>
Change all unquoted commas to a different character;
</para>
</listitem>

<listitem>
<para>
Remove all quotation marks.
</para>
</listitem>

</itemizedlist>
<para>
Strictly speaking, I could use <application>sed</application> to remove
the first line from the file, but doing so in my own program
was very easy, so I decided to do it and reduce the size of
the pipeline.
</para>

<para>
At any rate, writing a program like this took me about
20 minutes. Writing a program that extracts the 11th field
from the <acronym>CSV</acronym> file would take a lot longer,
and I could not reuse it to extract some other field from some
other database.
</para>

<para>
This time I decided to let it do a little more work than
a typical tutorial program would:
</para>

<itemizedlist>
<listitem>
<para>
It parses its command line for options;
</para>
</listitem>

<listitem>
<para>
It displays proper usage if it finds wrong arguments;
</para>
</listitem>

<listitem>
<para>
It produces meaningful error messages.
</para>
</listitem>

</itemizedlist>
<para>
Here is its usage message:
</para>

<screen>Usage: csv [-t&lt;delim&gt;] [-c&lt;comma&gt;] [-p] [-o &lt;outfile&gt;] [-i &lt;infile&gt;]</screen>

<para>
All parameters are optional, and can appear in any order.
</para>

<para>
The <parameter>-t</parameter> parameter declares what to replace
the commas with. The <constant>tab</constant> is the default here.
For example, <parameter>-t;</parameter> will replace all unquoted
commas with semicolons.
</para>

<para>
I did not need the <parameter>-c</parameter> option, but it may
come in handy in the future. It lets me declare that I want a
character other than a comma replaced with something else.
For example, <parameter>-c@</parameter> will replace all at signs
(useful if you want to split a list of email addresses
to their user names and domains).
</para>

<para>
The <parameter>-p</parameter> option preserves the first line, i.e.,
it does not delete it. By default, we delete the first
line because in a <acronym>CSV</acronym> file it contains the field
names rather than data.
</para>

<para>
The <parameter>-i</parameter> and <parameter>-o</parameter>
options let me specify the input and the output files. Defaults
are <filename>stdin</filename> and <filename>stdout</filename>,
so this is a regular Unix filter.
</para>

<para>
I made sure that both <parameter>-i filename</parameter> and
<parameter>-ifilename</parameter> are accepted. I also made
sure that only one input and one output files may be
specified.
</para>

<para>
To get the 11th field of each record, I can now do:
</para>

<screen>&prompt.user; <userinput>csv '-t;' <replaceable>data.csv</replaceable> | awk '-F;' '{print $11}'</userinput></screen>

<para>
The code stores the options (except for the file descriptors)
in <varname role="register">EDX</varname>: The comma in <varname role="register">DH</varname>, the new
separator in <varname role="register">DL</varname>, and the flag for
the <parameter>-p</parameter> option in the highest bit of
<varname role="register">EDX</varname>, so a check for its sign will give us a
quick decision what to do.
</para>

<para>
Here is the code:
</para>

<programlisting>
;;;;;;; csv.asm ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Convert a comma-separated file to a something-else separated file.
;
; Started:	31-May-2001
; Updated:	 1-Jun-2001
;
; Copyright (c) 2001 G. Adam Stanislav
; All rights reserved.
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

%include	'system.inc'

%define	BUFSIZE	2048

section	.data
fd.in	dd	stdin
fd.out	dd	stdout
usg	db	'Usage: csv [-t&lt;delim&gt;] [-c&lt;comma&gt;] [-p] [-o &lt;outfile&gt;] [-i &lt;infile&gt;]', 0Ah
usglen	equ	$-usg
iemsg	db	"csv: Can't open input file", 0Ah
iemlen	equ	$-iemsg
oemsg	db	"csv: Can't create output file", 0Ah
oemlen	equ	$-oemsg

section .bss
ibuffer	resb	BUFSIZE
obuffer	resb	BUFSIZE

section	.text
align 4
ierr:
	push	dword iemlen
	push	dword iemsg
	push	dword stderr
	sys.write
	push	dword 1		; return failure
	sys.exit

align 4
oerr:
	push	dword oemlen
	push	dword oemsg
	push	dword stderr
	sys.write
	push	dword 2
	sys.exit

align 4
usage:
	push	dword usglen
	push	dword usg
	push	dword stderr
	sys.write
	push	dword 3
	sys.exit

align 4
global	_start
_start:
	add	esp, byte 8	; discard argc and argv[0]
	mov	edx, (',' &lt;&lt; 8) | 9

.arg:
	pop	ecx
	or	ecx, ecx
	je	near .init		; no more arguments

	; ECX contains the pointer to an argument
	cmp	byte [ecx], '-'
	jne	usage

	inc	ecx
	mov	ax, [ecx]

.o:
	cmp	al, 'o'
	jne	.i

	; Make sure we are not asked for the output file twice
	cmp	dword [fd.out], stdout
	jne	usage

	; Find the path to output file - it is either at [ECX+1],
	; i.e., -ofile --
	; or in the next argument,
	; i.e., -o file

	inc	ecx
	or	ah, ah
	jne	.openoutput
	pop	ecx
	jecxz	usage

.openoutput:
	push	dword 420	; file mode (644 octal)
	push	dword 0200h | 0400h | 01h
	; O_CREAT | O_TRUNC | O_WRONLY
	push	ecx
	sys.open
	jc	near oerr

	add	esp, byte 12
	mov	[fd.out], eax
	jmp	short .arg

.i:
	cmp	al, 'i'
	jne	.p

	; Make sure we are not asked twice
	cmp	dword [fd.in], stdin
	jne	near usage

	; Find the path to the input file
	inc	ecx
	or	ah, ah
	jne	.openinput
	pop	ecx
	or	ecx, ecx
	je near usage

.openinput:
	push	dword 0		; O_RDONLY
	push	ecx
	sys.open
	jc	near ierr		; open failed

	add	esp, byte 8
	mov	[fd.in], eax
	jmp	.arg

.p:
	cmp	al, 'p'
	jne	.t
	or	ah, ah
	jne	near usage
	or	edx, 1 &lt;&lt; 31
	jmp	.arg

.t:
	cmp	al, 't'		; redefine output delimiter
	jne	.c
	or	ah, ah
	je	near usage
	mov	dl, ah
	jmp	.arg

.c:
	cmp	al, 'c'
	jne	near usage
	or	ah, ah
	je	near usage
	mov	dh, ah
	jmp	.arg

align 4
.init:
	sub	eax, eax
	sub	ebx, ebx
	sub	ecx, ecx
	mov	edi, obuffer

	; See if we are to preserve the first line
	or	edx, edx
	js	.loop

.firstline:
	; get rid of the first line
	call	getchar
	cmp	al, 0Ah
	jne	.firstline

.loop:
	; read a byte from stdin
	call	getchar

	; is it a comma (or whatever the user asked for)?
	cmp	al, dh
	jne	.quote

	; Replace the comma with a tab (or whatever the user wants)
	mov	al, dl

.put:
	call	putchar
	jmp	short .loop

.quote:
	cmp	al, '"'
	jne	.put

	; Print everything until you get another quote or EOL. If it
	; is a quote, skip it. If it is EOL, print it.
.qloop:
	call	getchar
	cmp	al, '"'
	je	.loop

	cmp	al, 0Ah
	je	.put

	call	putchar
	jmp	short .qloop

align 4
getchar:
	or	ebx, ebx
	jne	.fetch

	call	read

.fetch:
	lodsb
	dec	ebx
	ret

read:
	jecxz	.read
	call	write

.read:
	push	dword BUFSIZE
	mov	esi, ibuffer
	push	esi
	push	dword [fd.in]
	sys.read
	add	esp, byte 12
	mov	ebx, eax
	or	eax, eax
	je	.done
	sub	eax, eax
	ret

align 4
.done:
	call	write		; flush output buffer

	; close files
	push	dword [fd.in]
	sys.close

	push	dword [fd.out]
	sys.close

	; return success
	push	dword 0
	sys.exit

align 4
putchar:
	stosb
	inc	ecx
	cmp	ecx, BUFSIZE
	je	write
	ret

align 4
write:
	jecxz	.ret	; nothing to write
	sub	edi, ecx	; start of buffer
	push	ecx
	push	edi
	push	dword [fd.out]
	sys.write
	add	esp, byte 12
	sub	eax, eax
	sub	ecx, ecx	; buffer is empty now
.ret:
	ret
</programlisting>

<para>
Much of it is taken from <filename>hex.asm</filename> above. But there
is one important difference: I no longer call <function>write</function>
whenever I am outputing a line feed. Yet, the code can be
used interactively.
</para>

<para>
I have found a better solution for the interactive problem
since I first started writing this chapter. I wanted to
make sure each line is printed out separately only when needed.
After all, there is no need to flush out every line when used
non-interactively.
</para>

<para>
The new solution I use now is to call <function>write</function> every
time I find the input buffer empty. That way, when running in
the interactive mode, the program reads one line from the user's
keyboard, processes it, and sees its input buffer is empty. It
flushes its output and reads the next line.
</para>

<sect3 id="x86-buffered-dark-side">
<title>The Dark Side of Buffering</title>
<para>
This change prevents a mysterious lockup
in a very specific case. I refer to it as the
<emphasis>dark side of buffering</emphasis>, mostly
because it presents a danger that is not
quite obvious.
</para>

<para>
It is unlikely to happen with a program like the
<application>csv</application> above, so let us consider yet
another filter: In this case we expect our input
to be raw data representing color values, such as
the <emphasis>red</emphasis>, <emphasis>green</emphasis>, and
<emphasis>blue</emphasis> intensities of a pixel. Our
output will be the negative of our input.
</para>

<para>
Such a filter would be very simple to write.
Most of it would look just like all the other
filters we have written so far, so I am only
going to show you its inner loop:
</para>

<programlisting>
.loop:
	call	getchar
	not	al		; Create a negative
	call	putchar
	jmp	short .loop
</programlisting>
<para>
Because this filter works with raw data,
it is unlikely to be used interactively.
</para>

<para>
But it could be called by image manipulation software.
And, unless it calls <function>write</function> before each call
to <function>read</function>, chances are it will lock up.
</para>

<para>
Here is what might happen:
</para>

<procedure><step><para>
The image editor will load our filter using the
C function <function>popen()</function>.
</para>
</step>
<step><para>
It will read the first row of pixels from
a bitmap or pixmap.
</para>
</step>
<step><para>
It will write the first row of pixels to
the <emphasis>pipe</emphasis> leading to
the <varname>fd.in</varname> of our filter.
</para>
</step>
<step><para>
Our filter will read each pixel
from its input, turn it to a negative,
and write it to its output buffer.
</para>
</step>
<step><para>
Our filter will call <function>getchar</function>
to fetch the next pixel.
</para>
</step>
<step><para>
<function>getchar</function> will find an empty
input buffer, so it will call
<function>read</function>.
</para>
</step>
<step><para>
<function>read</function> will call the
<function role="syscall">SYS_read</function> system call.
</para>
</step>
<step><para>
The <emphasis>kernel</emphasis> will suspend
our filter until the image editor
sends more data to the pipe.
</para>
</step>
<step><para>
The image editor will read from the
other pipe, connected to the
<varname>fd.out</varname> of our filter so it can set the first row of the
output image <emphasis>before</emphasis>
it sends us the second row of the input.
</para>
</step>
<step><para>
The <emphasis>kernel</emphasis> suspends
the image editor until it receives
some output from our filter, so it
can pass it on to the image editor.
</para>
</step>
</procedure>
<para>
At this point our filter waits for the image
editor to send it more data to process, while
the image editor is waiting for our filter
to send it the result of the processing
of the first row. But the result sits in
our output buffer.
</para>

<para>
The filter and the image editor will continue
waiting for each other forever (or, at least,
until they are killed). Our software has just
entered a
<link linkend="secure-race-conditions">race condition</link>.
</para>

<para>
This problem does not exist if our filter flushes
its output buffer <emphasis>before</emphasis> asking the
<emphasis>kernel</emphasis> for more input data.
</para>

</sect3>

</sect2>

</sect1>

<sect1 id="x86-fpu">
<title>Using the <acronym>FPU</acronym></title>
<para>
Strangely enough, most of assembly language literature does not
even mention the existence of the <acronym>FPU</acronym>,
or <emphasis>floating point unit</emphasis>, let alone discuss
programming it.
</para>

<para>
Yet, never does assembly language shine more than when
we create highly optimized <acronym>FPU</acronym>
code by doing things that can be done <emphasis>only</emphasis> in assembly language.</para>

<sect2 id="x86-fpu-organization"><title>Organization of the <acronym>FPU</acronym></title>
<para>
The <acronym>FPU</acronym> consists of 8 80&ndash;bit floating&ndash;point registers.
These are organized in a stack fashion&mdash;you can
<function>push</function> a value on <acronym>TOS</acronym>
(<emphasis>top of stack</emphasis>) and you can
<function>pop</function> it.
</para>

<para>
That said, the assembly language op codes are not <function role="opcode">push</function>
and <function role="opcode">pop</function> because those are already taken.</para>

<para>
You can <function>push</function> a value on <acronym>TOS</acronym>
by using <function role="opcode">fld</function>, <function role="opcode">fild</function>,
and <function role="opcode">fbld</function>. Several other op codes
let you <function>push</function> many common
<emphasis>constants</emphasis>&mdash;such as <emphasis>pi</emphasis>&mdash;on
the <acronym>TOS</acronym>.
</para>

<para>
Similarly, you can <function>pop</function> a value by
using <function role="opcode">fst</function>, <function role="opcode">fstp</function>,
<function role="opcode">fist</function>, <function role="opcode">fistp</function>, and
<function role="opcode">fbstp</function>. Actually, only the op
codes that end with a <emphasis>p</emphasis> will
literally <function>pop</function> the value,
the rest will <function>store</function> it
somewhere else without removing it from
the <acronym>TOS</acronym>.
</para>

<para>
We can transfer the data between the
<acronym>TOS</acronym> and the computer memory either as
a 32&ndash;bit, 64&ndash;bit, or 80&ndash;bit <emphasis>real</emphasis>,
a 16&ndash;bit, 32&ndash;bit, or 64&ndash;bit <emphasis>integer</emphasis>,
or an 80&ndash;bit <emphasis>packed decimal</emphasis>.
</para>

<para>
The 80&ndash;bit <emphasis>packed decimal</emphasis> is
a special case of <emphasis>binary coded
decimal</emphasis> which is very convenient when
converting between the <acronym>ASCII</acronym>
representation of data and the internal
data of the <acronym>FPU</acronym>. It allows us to use
18 significant digits.
</para>

<para>
No matter how we represent data in the memory,
the <acronym>FPU</acronym> always stores it in the 80&ndash;bit
<emphasis>real</emphasis> format in its registers.
</para>

<para>
Its internal precision is at least 19 decimal
digits, so even if we choose to display results
as <acronym>ASCII</acronym> in the full
18&ndash;digit precision, we are still showing
correct results.
</para>

<para>
We can perform mathematical operations on the
<acronym>TOS</acronym>: We can calculate its
<emphasis>sine</emphasis>, we can <emphasis>scale</emphasis> it
(i.e., we can multiply or divide it by a power
of 2), we can calculate its base&ndash;2
<emphasis>logarithm</emphasis>, and many other things.
</para>

<para>
We can also <emphasis>multiply</emphasis> or
<emphasis>divide</emphasis> it by, <emphasis>add</emphasis>
it to, or <emphasis>subtract</emphasis> it from,
any of the <acronym>FPU</acronym> registers (including
itself).
</para>

<para>
The official Intel op code for the
<acronym>TOS</acronym> is <varname role="register">st</varname>, and
for the <emphasis>registers</emphasis>
<varname role="register">st(0)</varname>&ndash;<varname role="register">st(7)</varname>.
<varname role="register">st</varname> and <varname role="register">st(0)</varname>, then,
refer to the same register.
</para>

<para>
For whatever reasons, the original author of
<application>nasm</application> has decided to use
different op codes, namely
<varname role="register">st0</varname>&ndash;<varname role="register">st7</varname>.
In other words, there are no parentheses,
and the <acronym>TOS</acronym> is always
<varname role="register">st0</varname>, never just <function role="opcode">st</function>.
</para>

<sect3 id="x86-fpu-packed-decimal">
<title>The Packed Decimal Format</title>
<para>
The <emphasis>packed decimal</emphasis> format
uses 10 bytes (80 bits) of
memory to represent 18 digits. The
number represented there is always an
<emphasis>integer</emphasis>.
</para>

<tip>
<para>
You can use it to get decimal places
by multiplying the <acronym>TOS</acronym>
by a power of 10 first.
</para>
</tip>

<para>
The highest bit of the highest byte
(byte 9) is the <emphasis>sign bit</emphasis>:
If it is set, the number is <emphasis>negative</emphasis>,
otherwise, it is <emphasis>positive</emphasis>.
The rest of the bits of this byte are unused/ignored.
</para>

<para>
The remaining 9 bytes store the 18 digits
of the number: 2 digits per byte.</para>

<para>
The <emphasis>more significant digit</emphasis> is
stored in the high <emphasis>nibble</emphasis>
(4 bits), the <emphasis>less significant
digit</emphasis> in the low <emphasis>nibble</emphasis>.
</para>

<para>
That said, you might think that <constant>-1234567</constant>
would be stored in the memory like this (using
hexadecimal notation):
</para>

<programlisting>
80 00 00 00 00 00 01 23 45 67
</programlisting>
<para>
Alas it is not! As with everything else of Intel make,
even the <emphasis>packed decimal</emphasis> is
<emphasis>little&ndash;endian</emphasis>.</para>

<para>
That means our <constant>-1234567</constant>
is stored like this:
</para>

<programlisting>
67 45 23 01 00 00 00 00 00 80
</programlisting>
<para>
Remember that, or you will be pulling your hair out
in desperation!
</para>

<note>
<para>
The book to read&mdash;if you can find it&mdash;is Richard Startz'
<ulink url="http://www.int80h.org/cgi-bin/isbn?isbn=013246604X">8087/80287/80387
for the IBM PC &amp; Compatibles</ulink>.
Though it does seem to take the fact about the
little&ndash;endian storage of the <emphasis>packed
decimal</emphasis> for granted. I kid you not about the
desperation of trying to figure out what was wrong
with the filter I show below <emphasis>before</emphasis>
it occurred to me I should try the
little&ndash;endian order even for this type of data.
</para>
</note>

</sect3>

</sect2>

<sect2 id="x86-pinhole-photography">
<title>Excursion to Pinhole Photography</title>
<para>
To write meaningful software, we must not only
understand our programming tools, but also the
field we are creating software for.
</para>

<para>
Our next filter will help us whenever we want
to build a <emphasis>pinhole camera</emphasis>,
so, we need some background in <emphasis>pinhole
photography</emphasis> before we can continue.
</para>

<sect3 id="x86-camera">
<title>The Camera</title>
<para>
The easiest way to describe any camera ever built
is as some empty space enclosed in some
lightproof material, with a small hole in the
enclosure.
</para>

<para>
The enclosure is usually sturdy (e.g., a box),
though sometimes it is flexible (the bellows).
It is quite dark inside the camera. However, the
hole lets light rays in through a single point
(though in some cases there may be several).
These light rays form an image, a representation
of whatever is outside the camera, in front of the
hole.
</para>

<para>
If some light sensitive material (such as film)
is placed inside the camera, it can capture the
image.</para>

<para>
The hole often contains a <emphasis>lens</emphasis>, or
a lens assembly, often called the <emphasis>objective</emphasis>.
</para>

</sect3>

<sect3 id="x86-the-pinhole">
<title>The Pinhole</title>
<para>
But, strictly speaking, the lens is not necessary:
The original cameras did not use a lens but a
<emphasis>pinhole</emphasis>. Even today, <emphasis>pinholes</emphasis>
are used, both as a tool to study how cameras
work, and to achieve a special kind of image.
</para>

<para>
The image produced by the <emphasis>pinhole</emphasis>
is all equally sharp. Or <emphasis>blurred</emphasis>.
There is an ideal size for a pinhole: If it is
either larger or smaller, the image loses its
sharpness.</para>

</sect3>

<sect3 id="x86-focal-length">
<title>Focal Length</title>
<para>
This ideal pinhole diameter is a function
of the square root of <emphasis>focal
length</emphasis>, which is the distance of the
pinhole from the film.
</para>

<programlisting>
	D = PC * sqrt(FL)
</programlisting>
<para>
In here, <varname>D</varname> is the
ideal diameter of the pinhole,
<varname>FL</varname> is the focal length,
and <constant>PC</constant> is a pinhole
constant. According to Jay Bender,
its value is <constant>0.04</constant>, while
Kenneth Connors has determined it to
be <constant>0.037</constant>. Others have
proposed other values. Plus, this
value is for the daylight only: Other types
of light will require a different constant,
whose value can only be determined by
experimentation.
</para>

</sect3>

<sect3 id="x86-f-number">
<title>The F&ndash;Number</title>
<para>
The f&ndash;number is a very useful measure of
how much light reaches the film. A light
meter can determine that, for example,
to expose a film of specific sensitivity
with f5.6 may require the exposure to last
1/1000 sec.</para>

<para>
It does not matter whether it is a 35&ndash;mm
camera, or a 6x9cm camera, etc.
As long as we know the f&ndash;number, we can determine
the proper exposure.
</para>

<para>
The f&ndash;number is easy to calculate:
</para>

<programlisting>
	F = FL / D
</programlisting>
<para>
In other words, the f&ndash;number equals the focal
length divided by the diameter of the pinhole.
It also means a higher f&ndash;number either implies
a smaller pinhole or a larger focal distance,
or both. That, in turn, implies, the higher
the f&ndash;number, the longer the exposure has to be.
</para>

<para>
Furthermore, while pinhole diameter and focal
distance are one&ndash;dimensional measurements,
both, the film and the pinhole, are two&ndash;dimensional.
That means that
if you have measured the exposure at f&ndash;number
<varname>A</varname> as <varname>t</varname>, then the exposure
at f&ndash;number <varname>B</varname> is:</para>

<programlisting>
	t * (B / A)&#178;
</programlisting>
</sect3>

<sect3 id="x86-normalized-f-number">
<title>Normalized F&ndash;Number</title>
<para>
While many modern cameras can change the diameter
of their pinhole, and thus their f&ndash;number, quite
smoothly and gradually, such was not always the case.
</para>

<para>
To allow for different f&ndash;numbers, cameras typically
contained a metal plate with several holes of
different sizes drilled to them.
</para>

<para>
Their sizes were chosen according to the above
formula in such a way that the resultant f&ndash;number
was one of standard f&ndash;numbers used on all cameras
everywhere. For example, a very old Kodak Duaflex IV
camera in my possession has three such holes for
f&ndash;numbers 8, 11, and 16.
</para>

<para>
A more recently made camera may offer f&ndash;numbers of
2.8, 4, 5.6, 8, 11,
16, 22, and 32 (as well as others).
These numbers were not chosen arbitrarily: They all are
powers of the square root of 2, though they may
be rounded somewhat.
</para>

</sect3>

<sect3 id="x86-f-stop">
<title>The F&ndash;Stop</title>
<para>
A typical camera is designed in such a way that setting
any of the normalized f&ndash;numbers changes the feel of the
dial. It will naturally <emphasis>stop</emphasis> in that
position. Because of that, these positions of the dial
are called f&ndash;stops.</para>

<para>
Since the f&ndash;numbers at each stop are powers of the
square root of 2, moving the dial by 1
stop will double the amount of light required for
proper exposure. Moving it by 2 stops will
quadruple the required exposure. Moving the dial by
3 stops will require the increase in exposure
8 times, etc.
</para>

</sect3>

</sect2>

<sect2 id="x86-pinhole-software">
<title>Designing the Pinhole Software</title>
<para>
We are now ready to decide what exactly we want our
pinhole software to do.
</para>

<sect3 id="xpinhole-processing-input">
<title>Processing Program Input</title>
<para>
Since its main purpose is to help us design a working
pinhole camera, we will use the <emphasis>focal
length</emphasis> as the input to the program. This is something
we can determine without software: Proper focal length
is determined by the size of the film and by the need
to shoot "regular" pictures, wide angle pictures, or
telephoto pictures.
</para>

<para>
Most of the programs we have written so far worked with
individual characters, or bytes, as their input: The
<application>hex</application> program converted individual bytes
into a hexadecimal number, the <application>csv</application>
program either let a character through, or deleted it,
or changed it to a different character, etc.
</para>

<para>
One program, <application>ftuc</application> used the state machine
to consider at most two input bytes at a time.
</para>

<para>
But our <application>pinhole</application> program cannot just
work with individual characters, it has to deal with
larger syntactic units.
</para>

<para>
For example, if we want the program to calculate the
pinhole diameter (and other values we will discuss
later) at the focal lengths of <constant>100 mm</constant>,
<constant>150 mm</constant>, and <constant>210 mm</constant>, we may want
to enter something like this:</para>

<screen><userinput>100, 150, 210</userinput></screen>
<para>
Our program needs to consider more than a single byte of
input at a time. When it sees the first <constant>1</constant>,
it must understand it is seeing the first digit of a
decimal number. When it sees the <constant>0</constant> and
the other <constant>0</constant>, it must know it is seeing
more digits of the same number.
</para>

<para>
When it encounters the first comma, it must know it is
no longer receiving the digits of the first number.
It must be able to convert the digits of the first number
into the value of <constant>100</constant>. And the digits of the
second number into the value of <constant>150</constant>. And,
of course, the digits of the third number into the
numeric value of <constant>210</constant>.
</para>

<para>
We need to decide what delimiters to accept: Do the
input numbers have to be separated by a comma? If so,
how do we treat two numbers separated by something else?
</para>

<para>
Personally, I like to keep it simple. Something either
is a number, so I process it. Or it is not a number,
so I discard it. I don't like the computer complaining
about me typing in an extra character when it is
<emphasis>obvious</emphasis> that it is an extra character. Duh!
</para>

<para>
Plus, it allows me to break up the monotony of computing
and type in a query instead of just a number:
</para>

<screen><userinput>What is the best pinhole diameter for the focal length of 150?</userinput></screen>
<para>
There is no reason for the computer to spit out
a number of complaints:
</para>

<screen>Syntax error: What
Syntax error: is
Syntax error: the
Syntax error: best</screen>
<para>
Et cetera, et cetera, et cetera.</para>

<para>
Secondly, I like the <constant>#</constant> character to denote
the start of a comment which extends to the end of the
line. This does not take too much effort to code, and
lets me treat input files for my software as executable
scripts.
</para>

<para>
In our case, we also need to decide what units the
input should come in: We choose <emphasis>millimeters</emphasis>
because that is how most photographers measure
the focus length.
</para>

<para>
Finally, we need to decide whether to allow the use
of the decimal point (in which case we must also
consider the fact that much of the world uses a
decimal <emphasis>comma</emphasis>).</para>

<para>
In our case allowing for the decimal point/comma
would offer a false sense of precision: There is
little if any noticeable difference between the
focus lengths of <constant>50</constant> and <constant>51</constant>,
so allowing the user to input something like
<constant>50.5</constant> is not a good idea. This is
my opinion, mind you, but I am the one writing
this program. You can make other choices in yours,
of course.
</para>

</sect3>

<sect3 id="x86-pinhole-options">
<title>Offering Options</title>
<para>
The most important thing we need to know when building
a pinhole camera is the diameter of the pinhole. Since
we want to shoot sharp images, we will use the above
formula to calculate the pinhole diameter from focal length.
As experts are offering several different values for the
<constant>PC</constant> constant, we will need to have the choice.
</para>

<para>
It is traditional in Unix programming to have two main ways
of choosing program parameters, plus to have a default for
the time the user does not make a choice.
</para>

<para>
Why have two ways of choosing?</para>

<para>
One is to allow a (relatively) <emphasis>permanent</emphasis>
choice that applies automatically each time the
software is run without us having to tell it over and
over what we want it to do.
</para>

<para>
The permanent choices may be stored in a configuration
file, typically found in the user's home directory.
The file usually has the same name as the application
but is started with a dot. Often <emphasis>"rc"</emphasis>
is added to the file name. So, ours could be
<filename>~/.pinhole</filename> or <filename>~/.pinholerc</filename>.
(The <filename>~/</filename> means current user's
home directory.)
</para>

<para>
The configuration file is used mostly by programs
that have many configurable parameters. Those
that have only one (or a few) often use a different
method: They expect to find the parameter in an
<emphasis>environment variable</emphasis>. In our case,
we might look at an environment variable named
<varname>PINHOLE</varname>.
</para>

<para>
Usually, a program uses one or the other of the
above methods. Otherwise, if a configuration
file said one thing, but an environment variable
another, the program might get confused (or just
too complicated).
</para>

<para>
Because we only need to choose <emphasis>one</emphasis>
such parameter, we will go with the second method
and search the environment for a variable named
<varname>PINHOLE</varname>.</para>

<para>
The other way allows us to make <emphasis>ad hoc</emphasis>
decisions: <emphasis>"Though I usually want
you to use 0.039, this time I want 0.03872."</emphasis>
In other words, it allows us to <emphasis>override</emphasis>
the permanent choice.
</para>

<para>
This type of choice is usually done with command
line parameters.
</para>

<para>
Finally, a program <emphasis>always</emphasis> needs a
<emphasis>default</emphasis>. The user may not make
any choices. Perhaps he does not know what
to choose. Perhaps he is "just browsing."
Preferably, the default will be the value
most users would choose anyway. That way
they do not need to choose. Or, rather, they
can choose the default without an additional
effort.
</para>

<para>
Given this system, the program may find conflicting
options, and handle them this way:
</para>

<procedure><step><para>
If it finds an <emphasis>ad hoc</emphasis> choice
(e.g., command line parameter), it should
accept that choice. It must ignore any permanent
choice and any default.
</para>
</step>
<step><para>
<emphasis>Otherwise</emphasis>, if it finds
a permanent option (e.g., an environment
variable), it should accept it, and ignore
the default.</para>
</step>
<step><para>
<emphasis>Otherwise</emphasis>, it should use
the default.
</para>
</step>
</procedure>
<para>
We also need to decide what <emphasis>format</emphasis>
our <constant>PC</constant> option should have.
</para>

<para>
At first site, it seems obvious to use the
<varname>PINHOLE=0.04</varname> format for the
environment variable, and <parameter>-p0.04</parameter>
for the command line.
</para>

<para>
Allowing that is actually a security risk.
The <constant>PC</constant> constant is a very small
number. Naturally, we will test our software
using various small values of <constant>PC</constant>.
But what will happen if someone runs the program
choosing a huge value?
</para>

<para>
It may crash the program because we have not
designed it to handle huge numbers.
</para>

<para>
Or, we may spend more time on the program so
it can handle huge numbers. We might do that
if we were writing commercial software for
computer illiterate audience.
</para>

<para>
Or, we might say, <emphasis>"Tough!
The user should know better.""</emphasis>
</para>

<para>
Or, we just may make it impossible for the user
to enter a huge number. This is the approach we
will take: We will use an <emphasis>implied 0.</emphasis>
prefix.
</para>

<para>
In other words, if the user wants <constant>0.04</constant>,
we will expect him to type <parameter>-p04</parameter>,
or set <varname>PINHOLE=04</varname> in his environment.
So, if he says <parameter>-p9999999</parameter>, we will
interpret it as <constant>0.9999999</constant>&mdash;still
ridiculous but at least safer.
</para>

<para>
Secondly, many users will just want to go with either
Bender's constant or Connors' constant.
To make it easier on them, we will interpret
<parameter>-b</parameter> as identical to <parameter>-p04</parameter>,
and <parameter>-c</parameter> as identical to <parameter>-p037</parameter>.
</para>

</sect3>

<sect3 id="x86-pinhole-output">
<title>The Output</title>
<para>
We need to decide what we want our software to
send to the output, and in what format.
</para>

<para>
Since our input allows for an unspecified number
of focal length entries, it makes sense to use
a traditional database&ndash;style output of showing
the result of the calculation for each
focal length on a separate line, while
separating all values on one line by a
<constant>tab</constant> character.
</para>

<para>
Optionally, we should also allow the user
to specify the use of the <acronym>CSV</acronym>
format we have studied earlier. In this case,
we will print out a line of comma&ndash;separated
names describing each field of every line,
then show our results as before, but substituting
a <constant>comma</constant> for the <constant>tab</constant>.</para>

<para>
We need a command line option for the <acronym>CSV</acronym>
format. We cannot use <parameter>-c</parameter> because
that already means <emphasis>use Connors' constant</emphasis>.
For some strange reason, many web sites refer to
<acronym>CSV</acronym> files as <emphasis>"Excel
spreadsheet"</emphasis> (though the <acronym>CSV</acronym>
format predates Excel). We will, therefore, use
the <parameter>-e</parameter> switch to inform our software
we want the output in the <acronym>CSV</acronym> format.
</para>

<para>
We will start each line of the output with the
focal length. This may sound repetitious at first,
especially in the interactive mode: The user
types in the focal length, and we are repeating it.
</para>

<para>
But the user can type several focal lengths on one
line. The input can also come in from a file or
from the output of another program. In that case
the user does not see the input at all.
</para>

<para>
By the same token, the output can go to a file
which we will want to examine later, or it could
go to the printer, or become the input of another
program.
</para>

<para>
So, it makes perfect sense to start each line with
the focal length as entered by the user.
</para>

<para>
No, wait! Not as entered by the user. What if the user
types in something like this:</para>

<screen><userinput>00000000150</userinput></screen>
<para>
Clearly, we need to strip those leading zeros.</para>

<para>
So, we might consider reading the user input as is,
converting it to binary inside the <acronym>FPU</acronym>,
and printing it out from there.
</para>

<para>
But...</para>

<para>
What if the user types something like this:
</para>

<screen><userinput>17459765723452353453534535353530530534563507309676764423</userinput></screen>
<para>
Ha! The packed decimal <acronym>FPU</acronym> format
lets us input 18&ndash;digit numbers. But the
user has entered more than 18 digits. How
do we handle that?
</para>

<para>
Well, we <emphasis>could</emphasis> modify our code to read
the first 18 digits, enter it to the <acronym>FPU</acronym>,
then read more, multiply what we already have on the
<acronym>TOS</acronym> by 10 raised to the number
of additional digits, then <function>add</function> to it.
</para>

<para>
Yes, we could do that. But in <emphasis>this</emphasis>
program it would be ridiculous (in a different one it may be just the thing to do): Even the circumference of the Earth expressed in
millimeters only takes 11 digits. Clearly,
we cannot build a camera that large (not yet,
anyway).
</para>

<para>
So, if the user enters such a huge number, he is
either bored, or testing us, or trying to break
into the system, or playing games&mdash;doing
anything but designing a pinhole camera.
</para>

<para>
What will we do?</para>

<para>
We will slap him in the face, in a manner of speaking:</para>

<screen>17459765723452353453534535353530530534563507309676764423	???	???	???	???	???</screen>
<para>
To achieve that, we will simply ignore any leading zeros.
Once we find a non&ndash;zero digit, we will initialize a
counter to <constant>0</constant> and start taking three steps:
</para>

<procedure>
<step><para>
Send the digit to the output.
</para>
</step>
<step><para>
Append the digit to a buffer we will use later to
produce the packed decimal we can send to the
<acronym>FPU</acronym>.
</para>
</step>
<step><para>
Increase the counter.
</para>
</step>
</procedure>
<para>
Now, while we are taking these three steps,
we also need to watch out for one of two
conditions:</para>

<itemizedlist>
<listitem>
<para>
If the counter grows above 18,
we stop appending to the buffer. We
continue reading the digits and sending
them to the output.
</para>
</listitem>

<listitem>
<para>
If, or rather <emphasis>when</emphasis>,
the next input character is not
a digit, we are done inputting
for now.
</para>

<para>
Incidentally, we can simply
discard the non&ndash;digit, unless it
is a <constant>#</constant>, which we must
return to the input stream. It
starts a comment, so we must see it
after we are done producing output
and start looking for more input.
</para>
</listitem>

</itemizedlist>
<para>
That still leaves one possibility
uncovered: If all the user enters
is a zero (or several zeros), we
will never find a non&ndash;zero to
display.</para>

<para>
We can determine this has happened
whenever our counter stays at <constant>0</constant>.
In that case we need to send <constant>0</constant>
to the output, and perform another
"slap in the face":
</para>

<screen>0	???	???	???	???	???</screen>
<para>
Once we have displayed the focal
length and determined it is valid
(greater than <constant>0</constant>
but not exceeding 18 digits),
we can calculate the pinhole diameter.
</para>

<para>
It is not by coincidence that <emphasis>pinhole</emphasis>
contains the word <emphasis>pin</emphasis>. Indeed,
many a pinhole literally is a <emphasis>pin
hole</emphasis>, a hole carefully punched with the
tip of a pin.
</para>

<para>
That is because a typical pinhole is very
small. Our formula gets the result in
millimeters. We will multiply it by <constant>1000</constant>,
so we can output the result in <emphasis>microns</emphasis>.
</para>

<para>
At this point we have yet another trap to face:
<emphasis>Too much precision.</emphasis>
</para>

<para>
Yes, the <acronym>FPU</acronym> was designed
for high precision mathematics. But we
are not dealing with high precision
mathematics. We are dealing with physics
(optics, specifically).
</para>

<para>
Suppose we want to convert a truck into
a pinhole camera (we would not be the
first ones to do that!). Suppose its box is
<constant>12</constant>
meters long, so we have the focal length
of <constant>12000</constant>. Well, using Bender's constant, it gives us square root of
<constant>12000</constant> multiplied by <constant>0.04</constant>,
which is <constant>4.381780460</constant> millimeters,
or <constant>4381.780460</constant> microns.
</para>

<para>
Put either way, the result is absurdly precise.
Our truck is not <emphasis>exactly</emphasis> <constant>12000</constant>
millimeters long. We did not measure its length
with such a precision, so stating we need a pinhole
with the diameter of <constant>4.381780460</constant>
millimeters is, well, deceiving. <constant>4.4</constant>
millimeters would do just fine.
</para>

<note>
<para>
I "only" used ten digits in the above example.
Imagine the absurdity of going for all 18!
</para>
</note>

<para>
We need to limit the number of significant
digits of our result. One way of doing it
is by using an integer representing microns.
So, our truck would need a pinhole with the diameter
of <constant>4382</constant> microns. Looking at that number, we still decide that <constant>4400</constant> microns,
or <constant>4.4</constant> millimeters is close enough.
</para>

<para>
Additionally, we can decide that no matter how
big a result we get, we only want to display four
siginificant digits (or any other number
of them, of course). Alas, the <acronym>FPU</acronym>
does not offer rounding to a specific number
of digits (after all, it does not view the
numbers as decimal but as binary).
</para>

<para>
We, therefore, must devise an algorithm to reduce
the number of significant digits.
</para>

<para>
Here is mine (I think it is awkward&mdash;if
you know a better one, <emphasis>please</emphasis>, let me know):</para>

<procedure>
<step><para>
Initialize a counter to <constant>0</constant>.
</para>
</step>
<step><para>
While the number is greater than or equal to
<constant>10000</constant>, divide it by
<constant>10</constant> and increase the counter.
</para>
</step>
<step><para>
Output the result.</para>
</step>
<step><para>
While the counter is greater than <constant>0</constant>,
output <constant>0</constant> and decrease the counter.
</para>
</step>
</procedure>
<note>
<para>
The <constant>10000</constant> is only good if you want
<emphasis>four</emphasis> significant digits. For any other
number of significant digits, replace
<constant>10000</constant> with <constant>10</constant>
raised to the number of significant digits.
</para>
</note>

<para>
We will, then, output the pinhole diameter
in microns, rounded off to four significant
digits.
</para>

<para>
At this point, we know the <emphasis>focal
length</emphasis> and the <emphasis>pinhole
diameter</emphasis>. That means we have enough
information to also calculate the
<emphasis>f&ndash;number</emphasis>.
</para>

<para>
We will display the f&ndash;number, rounded to
four significant digits. Chances are the
f&ndash;number will tell us very little. To make
it more meaningful, we can find the nearest
<emphasis>normalized f&ndash;number</emphasis>, i.e.,
the nearest power of the square root
of 2.
</para>

<para>
We do that by multiplying the actual f&ndash;number
by itself, which, of course, will give us
its <function>square</function>. We will then calculate
its base&ndash;2 logarithm, which is much
easier to do than calculating the
base&ndash;square&ndash;root&ndash;of&ndash;2 logarithm!
We will round the result to the nearest integer.
Next, we will raise 2 to the result. Actually,
the <acronym>FPU</acronym> gives us a good shortcut
to do that: We can use the <function role="opcode">fscale</function>
op code to "scale" 1, which is
analogous to <function role="opcode">shift</function>ing an
integer left. Finally, we calculate the square
root of it all, and we have the nearest
normalized f&ndash;number.
</para>

<para>
If all that sounds overwhelming&mdash;or too much
work, perhaps&mdash;it may become much clearer
if you see the code. It takes 9 op
codes altogether:</para>

<programlisting>
	fmul	st0, st0
	fld1
	fld	st1
	fyl2x
	frndint
	fld1
	fscale
	fsqrt
	fstp	st1
</programlisting>
<para>
The first line, <function role="opcode">fmul st0, st0</function>, squares
the contents of the <acronym>TOS</acronym>
(top of the stack, same as <varname role="register">st</varname>,
called <varname role="register">st0</varname> by <application>nasm</application>).
The <function role="opcode">fld1</function> pushes <constant>1</constant>
on the <acronym>TOS</acronym>.</para>

<para>
The next line, <function role="opcode">fld st1</function>, pushes
the square back to the <acronym>TOS</acronym>.
At this point the square is both in <varname role="register">st</varname>
and <varname role="register">st(2)</varname> (it will become
clear why we leave a second copy on the stack
in a moment). <varname role="register">st(1)</varname> contains
<constant>1</constant>.
</para>

<para>
Next, <function role="opcode">fyl2x</function> calculates base&ndash;2
logarithm of <varname role="register">st</varname> multiplied by
<varname role="register">st(1)</varname>. That is why we placed <constant>1</constant> on <varname role="register">st(1)</varname> before.</para>

<para>
At this point, <varname role="register">st</varname> contains
the logarithm we have just calculated,
<varname role="register">st(1)</varname> contains the square
of the actual f&ndash;number we saved for later.
</para>

<para>
<function role="opcode">frndint</function> rounds the <acronym>TOS</acronym>
to the nearest integer. <function role="opcode">fld1</function> pushes
a <constant>1</constant>. <function role="opcode">fscale</function> shifts the
<constant>1</constant> we have on the <acronym>TOS</acronym>
by the value in <varname role="register">st(1)</varname>,
effectively raising 2 to <varname role="register">st(1)</varname>.
</para>

<para>
Finally, <function role="opcode">fsqrt</function> calculates
the square root of the result, i.e.,
the nearest normalized f&ndash;number.
</para>

<para>
We now have the nearest normalized
f&ndash;number on the <acronym>TOS</acronym>,
the base&ndash;2 logarithm rounded to the
nearest integer in <varname role="register">st(1)</varname>,
and the square of the actual f&ndash;number
in <varname role="register">st(2)</varname>. We are saving
the value in <varname role="register">st(2)</varname> for later.
</para>

<para>
But we do not need the contents of
<varname role="register">st(1)</varname> anymore. The last
line, <function role="opcode">fstp st1</function>, places the
contents of <varname role="register">st</varname> to
<varname role="register">st(1)</varname>, and pops. As a
result, what was <varname role="register">st(1)</varname>
is now <varname role="register">st</varname>, what was <varname role="register">st(2)</varname>
is now <varname role="register">st(1)</varname>, etc.
The new <varname role="register">st</varname> contains the
normalized f&ndash;number. The new
<varname role="register">st(1)</varname> contains the square
of the actual f&ndash;number we have
stored there for posterity.
</para>

<para>
At this point, we are ready to output
the normalized f&ndash;number. Because it is
normalized, we will not round it off to
four significant digits, but will
send it out in its full precision.
</para>

<para>
The normalized f-number is useful as long
as it is reasonably small and can be found
on our light meter. Otherwise we need a
different method of determining proper
exposure.
</para>

<para>
Earlier we have figured out the formula
of calculating proper exposure at an arbitrary
f&ndash;number from that measured at a different
f&ndash;number.
</para>

<para>
Every light meter I have ever seen can determine
proper exposure at f5.6. We will, therefore,
calculate an <emphasis>"f5.6 multiplier,"</emphasis>
i.e., by how much we need to multiply the exposure measured
at f5.6 to determine the proper exposure
for our pinhole camera.
</para>

<para>
From the above formula we know this factor can be
calculated by dividing our f&ndash;number (the
actual one, not the normalized one) by
<constant>5.6</constant>, and squaring the result.
</para>

<para>
Mathematically, dividing the square of our
f&ndash;number by the square of <constant>5.6</constant>
will give us the same result.
</para>

<para>
Computationally, we do not want to square
two numbers when we can only square one.
So, the first solution seems better at first.
</para>

<para>
But...</para>

<para>
<constant>5.6</constant> is a <emphasis>constant</emphasis>.
We do not have to have our <acronym>FPU</acronym>
waste precious cycles. We can just tell it
to divide the square of the f&ndash;number by
whatever <constant>5.6&#178;</constant> equals to.
Or we can divide the f&ndash;number by <constant>5.6</constant>,
and then square the result. The two ways
now seem equal.
</para>

<para>
But, they are not!</para>

<para>
Having studied the principles of photography
above, we remember that the <constant>5.6</constant>
is actually square root of 2 raised to
the fifth power. An <emphasis>irrational</emphasis>
number. The square of this number is
<emphasis>exactly</emphasis> <constant>32</constant>.
</para>

<para>
Not only is <constant>32</constant> an integer,
it is a power of 2. We do not need
to divide the square of the f&ndash;number by
<constant>32</constant>. We only need to use
<function role="opcode">fscale</function> to shift it right by
five positions. In the <acronym>FPU</acronym>
lingo it means we will <function role="opcode">fscale</function> it
with <varname role="register">st(1)</varname> equal to
<constant>-5</constant>. That is <emphasis>much
faster</emphasis> than a division.
</para>

<para>
So, now it has become clear why we have
saved the square of the f&ndash;number on the
top of the <acronym>FPU</acronym> stack.
The calculation of the f5.6 multiplier
is the easiest calculation of this
entire program! We will output it rounded
to four significant digits.
</para>

<para>
There is one more useful number we can calculate:
The number of stops our f&ndash;number is from f5.6.
This may help us if our f&ndash;number is just outside
the range of our light meter, but we have
a shutter which lets us set various speeds,
and this shutter uses stops.
</para>

<para>
Say, our f&ndash;number is 5 stops from
f5.6, and the light meter says
we should use 1/1000 sec.
Then we can set our shutter speed to 1/1000
first, then move the dial by 5 stops.
</para>

<para>
This calculation is quite easy as well. All
we have to do is to calculate the base-2
logarithm of the f5.6 multiplier
we had just calculated (though we need its
value from before we rounded it off). We then
output the result rounded to the nearest integer.
We do not need to worry about having more than
four significant digits in this one: The result
is most likely to have only one or two digits
anyway.</para>

</sect3>

</sect2>

<sect2 id="x86-fpu-optimizations">
<title>FPU Optimizations</title>
<para>
In assembly language we can optimize the <acronym>FPU</acronym>
code in ways impossible in high languages,
including C.
</para>

<para>
Whenever a C function needs to calculate
a floating&ndash;point value, it loads all necessary
variables and constants into <acronym>FPU</acronym>
registers. It then does whatever calculation is
required to get the correct result. Good C
compilers can optimize that part of the code really
well.
</para>

<para>
It "returns" the value by leaving
the result on the <acronym>TOS</acronym>.
However, before it returns, it cleans up.
Any variables and constants it used in its
calculation are now gone from the <acronym>FPU</acronym>.
</para>

<para>
It cannot do what we just did above: We calculated
the square of the f&ndash;number and kept it on the
stack for later use by another function.
</para>

<para>
We <emphasis>knew</emphasis> we would need that value
later on. We also knew we had enough room on the
stack (which only has room for 8 numbers)
to store it there.
</para>

<para>
A C compiler has no way of knowing
that a value it has on the stack will be
required again in the very near future.
</para>

<para>
Of course, the C programmer may know it.
But the only recourse he has is to store the
value in a memory variable.
</para>

<para>
That means, for one, the value will be changed
from the 80-bit precision used internally
by the <acronym>FPU</acronym> to a C <emphasis>double</emphasis>
(64 bits) or even <emphasis>single</emphasis> (32
bits).
</para>

<para>
That also means that the value must be moved
from the <acronym>TOS</acronym> into the memory,
and then back again. Alas, of all <acronym>FPU</acronym>
operations, the ones that access the computer
memory are the slowest.
</para>

<para>
So, whenever programming the <acronym>FPU</acronym>
in assembly language, look for the ways of keeping
intermediate results on the <acronym>FPU</acronym>
stack.
</para>

<para>
We can take that idea even further! In our
program we are using a <emphasis>constant</emphasis>
(the one we named <constant>PC</constant>).
</para>

<para>
It does not matter how many pinhole diameters
we are calculating: 1, 10, 20,
1000, we are always using the same constant.
Therefore, we can optimize our program by keeping
the constant on the stack all the time.
</para>

<para>
Early on in our program, we are calculating the
value of the above constant. We need to divide
our input by <constant>10</constant> for every digit in the
constant.
</para>

<para>
It is much faster to multiply than to divide.
So, at the start of our program, we divide <constant>10</constant>
into <constant>1</constant> to obtain <constant>0.1</constant>, which we
then keep on the stack: Instead of dividing the
input by <constant>10</constant> for every digit,
we multiply it by <constant>0.1</constant>.
</para>

<para>
By the way, we do not input <constant>0.1</constant> directly,
even though we could. We have a reason for that:
While <constant>0.1</constant> can be expressed with just one
decimal place, we do not know how many <emphasis>binary</emphasis>
places it takes. We, therefore, let the <acronym>FPU</acronym>
calculate its binary value to its own high precision.
</para>

<para>
We are using other constants: We multiply the pinhole
diameter by <constant>1000</constant> to convert it from
millimeters to microns. We compare numbers to
<constant>10000</constant> when we are rounding them off to
four significant digits. So, we keep both, <constant>1000</constant>
and <constant>10000</constant>, on the stack. And, of course,
we reuse the <constant>0.1</constant> when rounding off numbers
to four digits.
</para>

<para>
Last but not least, we keep <constant>-5</constant> on the stack.
We need it to scale the square of the f&ndash;number,
instead of dividing it by <constant>32</constant>. It is not
by coincidence we load this constant last. That makes
it the top of the stack when only the constants
are on it. So, when the square of the f&ndash;number is
being scaled, the <constant>-5</constant> is at <varname role="register">st(1)</varname>,
precisely where <function role="opcode">fscale</function> expects it to be.
</para>

<para>
It is common to create certain constants from
scratch instead of loading them from the memory.
That is what we are doing with <constant>-5</constant>:
</para>

<programlisting>
	fld1			; TOS =  1
	fadd	st0, st0	; TOS =  2
	fadd	st0, st0	; TOS =  4
	fld1			; TOS =  1
	faddp	st1, st0	; TOS =  5
	fchs			; TOS = -5
</programlisting>
<para>
We can generalize all these optimizations into one rule:
<emphasis>Keep repeat values on the stack!</emphasis>
</para>

<tip>
<para>
<emphasis>PostScript</emphasis> is a stack&ndash;oriented
programming language. There are many more books
available about PostScript than about the
<acronym>FPU</acronym> assembly language: Mastering
PostScript will help you master the <acronym>FPU</acronym>.
</para>
</tip>

</sect2>

<sect2 id="x86-pinhole-the-code">
<title><application>pinhole</application>&mdash;The Code</title>
<programlisting>
;;;;;;; pinhole.asm ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;
; Find various parameters of a pinhole camera construction and use
;
; Started:	 9-Jun-2001
; Updated:	10-Jun-2001
;
; Copyright (c) 2001 G. Adam Stanislav
; All rights reserved.
;
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

%include	'system.inc'

%define	BUFSIZE	2048

section	.data
align 4
ten	dd	10
thousand	dd	1000
tthou	dd	10000
fd.in	dd	stdin
fd.out	dd	stdout
envar	db	'PINHOLE='	; Exactly 8 bytes, or 2 dwords long
pinhole	db	'04,', 		; Bender's constant (0.04)
connors	db	'037', 0Ah	; Connors' constant
usg	db	'Usage: pinhole [-b] [-c] [-e] [-p &lt;value&gt;] [-o &lt;outfile&gt;] [-i &lt;infile&gt;]', 0Ah
usglen	equ	$-usg
iemsg	db	"pinhole: Can't open input file", 0Ah
iemlen	equ	$-iemsg
oemsg	db	"pinhole: Can't create output file", 0Ah
oemlen	equ	$-oemsg
pinmsg	db	"pinhole: The PINHOLE constant must not be 0", 0Ah
pinlen	equ	$-pinmsg
toobig	db	"pinhole: The PINHOLE constant may not exceed 18 decimal places", 0Ah
biglen	equ	$-toobig
huhmsg	db	9, '???'
separ	db	9, '???'
sep2	db	9, '???'
sep3	db	9, '???'
sep4	db	9, '???', 0Ah
huhlen	equ	$-huhmsg
header	db	'focal length in millimeters,pinhole diameter in microns,'
	db	'F-number,normalized F-number,F-5.6 multiplier,stops '
	db	'from F-5.6', 0Ah
headlen	equ	$-header

section .bss
ibuffer	resb	BUFSIZE
obuffer	resb	BUFSIZE
dbuffer	resb	20		; decimal input buffer
bbuffer	resb	10		; BCD buffer

section	.text
align 4
huh:
	call	write
	push	dword huhlen
	push	dword huhmsg
	push	dword [fd.out]
	sys.write
	add	esp, byte 12
	ret

align 4
perr:
	push	dword pinlen
	push	dword pinmsg
	push	dword stderr
	sys.write
	push	dword 4		; return failure
	sys.exit

align 4
consttoobig:
	push	dword biglen
	push	dword toobig
	push	dword stderr
	sys.write
	push	dword 5		; return failure
	sys.exit

align 4
ierr:
	push	dword iemlen
	push	dword iemsg
	push	dword stderr
	sys.write
	push	dword 1		; return failure
	sys.exit

align 4
oerr:
	push	dword oemlen
	push	dword oemsg
	push	dword stderr
	sys.write
	push	dword 2
	sys.exit

align 4
usage:
	push	dword usglen
	push	dword usg
	push	dword stderr
	sys.write
	push	dword 3
	sys.exit

align 4
global	_start
_start:
	add	esp, byte 8	; discard argc and argv[0]
	sub	esi, esi

.arg:
	pop	ecx
	or	ecx, ecx
	je	near .getenv		; no more arguments

	; ECX contains the pointer to an argument
	cmp	byte [ecx], '-'
	jne	usage

	inc	ecx
	mov	ax, [ecx]
	inc	ecx

.o:
	cmp	al, 'o'
	jne	.i

	; Make sure we are not asked for the output file twice
	cmp	dword [fd.out], stdout
	jne	usage

	; Find the path to output file - it is either at [ECX+1],
	; i.e., -ofile --
	; or in the next argument,
	; i.e., -o file

	or	ah, ah
	jne	.openoutput
	pop	ecx
	jecxz	usage

.openoutput:
	push	dword 420	; file mode (644 octal)
	push	dword 0200h | 0400h | 01h
	; O_CREAT | O_TRUNC | O_WRONLY
	push	ecx
	sys.open
	jc	near oerr

	add	esp, byte 12
	mov	[fd.out], eax
	jmp	short .arg

.i:
	cmp	al, 'i'
	jne	.p

	; Make sure we are not asked twice
	cmp	dword [fd.in], stdin
	jne	near usage

	; Find the path to the input file
	or	ah, ah
	jne	.openinput
	pop	ecx
	or	ecx, ecx
	je near usage

.openinput:
	push	dword 0		; O_RDONLY
	push	ecx
	sys.open
	jc	near ierr		; open failed

	add	esp, byte 8
	mov	[fd.in], eax
	jmp	.arg

.p:
	cmp	al, 'p'
	jne	.c
	or	ah, ah
	jne	.pcheck

	pop	ecx
	or	ecx, ecx
	je	near usage

	mov	ah, [ecx]

.pcheck:
	cmp	ah, '0'
	jl	near usage
	cmp	ah, '9'
	ja	near usage
	mov	esi, ecx
	jmp	.arg

.c:
	cmp	al, 'c'
	jne	.b
	or	ah, ah
	jne	near usage
	mov	esi, connors
	jmp	.arg

.b:
	cmp	al, 'b'
	jne	.e
	or	ah, ah
	jne	near usage
	mov	esi, pinhole
	jmp	.arg

.e:
	cmp	al, 'e'
	jne	near usage
	or	ah, ah
	jne	near usage
	mov	al, ','
	mov	[huhmsg], al
	mov	[separ], al
	mov	[sep2], al
	mov	[sep3], al
	mov	[sep4], al
	jmp	.arg

align 4
.getenv:
	; If ESI = 0, we did not have a -p argument,
	; and need to check the environment for "PINHOLE="
	or	esi, esi
	jne	.init

	sub	ecx, ecx

.nextenv:
	pop	esi
	or	esi, esi
	je	.default	; no PINHOLE envar found

	; check if this envar starts with 'PINHOLE='
	mov	edi, envar
	mov	cl, 2		; 'PINHOLE=' is 2 dwords long
rep	cmpsd
	jne	.nextenv

	; Check if it is followed by a digit
	mov	al, [esi]
	cmp	al, '0'
	jl	.default
	cmp	al, '9'
	jbe	.init
	; fall through

align 4
.default:
	; We got here because we had no -p argument,
	; and did not find the PINHOLE envar.
	mov	esi, pinhole
	; fall through

align 4
.init:
	sub	eax, eax
	sub	ebx, ebx
	sub	ecx, ecx
	sub	edx, edx
	mov	edi, dbuffer+1
	mov	byte [dbuffer], '0'

	; Convert the pinhole constant to real
.constloop:
	lodsb
	cmp	al, '9'
	ja	.setconst
	cmp	al, '0'
	je	.processconst
	jb	.setconst

	inc	dl

.processconst:
	inc	cl
	cmp	cl, 18
	ja	near consttoobig
	stosb
	jmp	short .constloop

align 4
.setconst:
	or	dl, dl
	je	near perr

	finit
	fild	dword [tthou]

	fld1
	fild	dword [ten]
	fdivp	st1, st0

	fild	dword [thousand]
	mov	edi, obuffer

	mov	ebp, ecx
	call	bcdload

.constdiv:
	fmul	st0, st2
	loop	.constdiv

	fld1
	fadd	st0, st0
	fadd	st0, st0
	fld1
	faddp	st1, st0
	fchs

	; If we are creating a CSV file,
	; print header
	cmp	byte [separ], ','
	jne	.bigloop

	push	dword headlen
	push	dword header
	push	dword [fd.out]
	sys.write

.bigloop:
	call	getchar
	jc	near done

	; Skip to the end of the line if you got '#'
	cmp	al, '#'
	jne	.num
	call	skiptoeol
	jmp	short .bigloop

.num:
	; See if you got a number
	cmp	al, '0'
	jl	.bigloop
	cmp	al, '9'
	ja	.bigloop

	; Yes, we have a number
	sub	ebp, ebp
	sub	edx, edx

.number:
	cmp	al, '0'
	je	.number0
	mov	dl, 1

.number0:
	or	dl, dl		; Skip leading 0's
	je	.nextnumber
	push	eax
	call	putchar
	pop	eax
	inc	ebp
	cmp	ebp, 19
	jae	.nextnumber
	mov	[dbuffer+ebp], al

.nextnumber:
	call	getchar
	jc	.work
	cmp	al, '#'
	je	.ungetc
	cmp	al, '0'
	jl	.work
	cmp	al, '9'
	ja	.work
	jmp	short .number

.ungetc:
	dec	esi
	inc	ebx

.work:
	; Now, do all the work
	or	dl, dl
	je	near .work0

	cmp	ebp, 19
	jae	near .toobig

	call	bcdload

	; Calculate pinhole diameter

	fld	st0	; save it
	fsqrt
	fmul	st0, st3
	fld	st0
	fmul	st5
	sub	ebp, ebp

	; Round off to 4 significant digits
.diameter:
	fcom	st0, st7
	fstsw	ax
	sahf
	jb	.printdiameter
	fmul	st0, st6
	inc	ebp
	jmp	short .diameter

.printdiameter:
	call	printnumber	; pinhole diameter

	; Calculate F-number

	fdivp	st1, st0
	fld	st0

	sub	ebp, ebp

.fnumber:
	fcom	st0, st6
	fstsw	ax
	sahf
	jb	.printfnumber
	fmul	st0, st5
	inc	ebp
	jmp	short .fnumber

.printfnumber:
	call	printnumber	; F number

	; Calculate normalized F-number
	fmul	st0, st0
	fld1
	fld	st1
	fyl2x
	frndint
	fld1
	fscale
	fsqrt
	fstp	st1

	sub	ebp, ebp
	call	printnumber

	; Calculate time multiplier from F-5.6

	fscale
	fld	st0

	; Round off to 4 significant digits
.fmul:
	fcom	st0, st6
	fstsw	ax
	sahf

	jb	.printfmul
	inc	ebp
	fmul	st0, st5
	jmp	short .fmul

.printfmul:
	call	printnumber	; F multiplier

	; Calculate F-stops from 5.6

	fld1
	fxch	st1
	fyl2x

	sub	ebp, ebp
	call	printnumber

	mov	al, 0Ah
	call	putchar
	jmp	.bigloop

.work0:
	mov	al, '0'
	call	putchar

align 4
.toobig:
	call	huh
	jmp	.bigloop

align 4
done:
	call	write		; flush output buffer

	; close files
	push	dword [fd.in]
	sys.close

	push	dword [fd.out]
	sys.close

	finit

	; return success
	push	dword 0
	sys.exit

align 4
skiptoeol:
	; Keep reading until you come to cr, lf, or eof
	call	getchar
	jc	done
	cmp	al, 0Ah
	jne	.cr
	ret

.cr:
	cmp	al, 0Dh
	jne	skiptoeol
	ret

align 4
getchar:
	or	ebx, ebx
	jne	.fetch

	call	read

.fetch:
	lodsb
	dec	ebx
	clc
	ret

read:
	jecxz	.read
	call	write

.read:
	push	dword BUFSIZE
	mov	esi, ibuffer
	push	esi
	push	dword [fd.in]
	sys.read
	add	esp, byte 12
	mov	ebx, eax
	or	eax, eax
	je	.empty
	sub	eax, eax
	ret

align 4
.empty:
	add	esp, byte 4
	stc
	ret

align 4
putchar:
	stosb
	inc	ecx
	cmp	ecx, BUFSIZE
	je	write
	ret

align 4
write:
	jecxz	.ret	; nothing to write
	sub	edi, ecx	; start of buffer
	push	ecx
	push	edi
	push	dword [fd.out]
	sys.write
	add	esp, byte 12
	sub	eax, eax
	sub	ecx, ecx	; buffer is empty now
.ret:
	ret

align 4
bcdload:
	; EBP contains the number of chars in dbuffer
	push	ecx
	push	esi
	push	edi

	lea	ecx, [ebp+1]
	lea	esi, [dbuffer+ebp-1]
	shr	ecx, 1

	std

	mov	edi, bbuffer
	sub	eax, eax
	mov	[edi], eax
	mov	[edi+4], eax
	mov	[edi+2], ax

.loop:
	lodsw
	sub	ax, 3030h
	shl	al, 4
	or	al, ah
	mov	[edi], al
	inc	edi
	loop	.loop

	fbld	[bbuffer]

	cld
	pop	edi
	pop	esi
	pop	ecx
	sub	eax, eax
	ret

align 4
printnumber:
	push	ebp
	mov	al, [separ]
	call	putchar

	; Print the integer at the TOS
	mov	ebp, bbuffer+9
	fbstp	[bbuffer]

	; Check the sign
	mov	al, [ebp]
	dec	ebp
	or	al, al
	jns	.leading

	; We got a negative number (should never happen)
	mov	al, '-'
	call	putchar

.leading:
	; Skip leading zeros
	mov	al, [ebp]
	dec	ebp
	or	al, al
	jne	.first
	cmp	ebp, bbuffer
	jae	.leading

	; We are here because the result was 0.
	; Print '0' and return
	mov	al, '0'
	jmp	putchar

.first:
	; We have found the first non-zero.
	; But it is still packed
	test	al, 0F0h
	jz	.second
	push	eax
	shr	al, 4
	add	al, '0'
	call	putchar
	pop	eax
	and	al, 0Fh

.second:
	add	al, '0'
	call	putchar

.next:
	cmp	ebp, bbuffer
	jb	.done

	mov	al, [ebp]
	push	eax
	shr	al, 4
	add	al, '0'
	call	putchar
	pop	eax
	and	al, 0Fh
	add	al, '0'
	call	putchar

	dec	ebp
	jmp	short .next

.done:
	pop	ebp
	or	ebp, ebp
	je	.ret

.zeros:
	mov	al, '0'
	call	putchar
	dec	ebp
	jne	.zeros

.ret:
	ret
</programlisting>
<para>
The code follows the same format as all the other
filters we have seen before, with one subtle
exception:
</para>

<blockquote>
<para>
We are no longer assuming that the end of input
implies the end of things to do, something we
took for granted in the <emphasis>character&ndash;oriented</emphasis>
filters.
</para>

<para>
This filter does not process characters. It
processes a <emphasis>language</emphasis>
(albeit a very simple
one, consisting only of numbers).
</para>

<para>
When we have no more input, it can mean one
of two things:</para>

<itemizedlist><listitem>
<para>
We are done and can quit. This is the
same as before.
</para>
</listitem>

<listitem>
<para>
The last character we have read was a digit.
We have stored it at the end of our
<acronym>ASCII</acronym>&ndash;to&ndash;float conversion
buffer. We now need to convert
the contents of that buffer into a
number and write the last line of our
output.
</para>
</listitem>

</itemizedlist>
<para>
For that reason, we have modified our <function>getchar</function>
and our <function>read</function> routines to return with
the <varname role="register">carry flag</varname> <emphasis>clear</emphasis> whenever we are
fetching another character from the input, or the
<varname role="register">carry flag</varname> <emphasis>set</emphasis> whenever there is no more
input.
</para>

<para>
Of course, we are still using assembly language magic
to do that! Take a good look at <function>getchar</function>.
It <emphasis>always</emphasis> returns with the
<varname role="register">carry flag</varname> <emphasis>clear</emphasis>.
</para>

<para>
Yet, our main code relies on the <varname role="register">carry
flag</varname> to tell it when to quit&mdash;and it works.
</para>

<para>
The magic is in <function>read</function>. Whenever it
receives more input from the system, it just
returns to <function>getchar</function>, which
fetches a character from the input buffer,
<emphasis>clears</emphasis> the <varname role="register">carry flag</varname>
and returns.
</para>

<para>
But when <function>read</function> receives no more
input from the system, it does <emphasis>not</emphasis>
return to <function>getchar</function> at all.
Instead, the <function role="opcode">add esp, byte 4</function>
op code adds <constant>4</constant> to <varname role="register">ESP</varname>,
<emphasis>sets</emphasis> the <varname role="register">carry
flag</varname>, and returns.
</para>

<para>
So, where does it return to? Whenever a
program uses the <function role="opcode">call</function> op code,
the microprocessor <function role="opcode">push</function>es the
return address, i.e., it stores it on
the top of the stack (not the <acronym>FPU</acronym>
stack, the system stack, which is in the memory).
When a program uses the <function role="opcode">ret</function>
op code, the microprocessor <function role="opcode">pop</function>s
the return value from the stack, and jumps
to the address that was stored there.
</para>

<para>
But since we added <constant>4</constant> to
<varname role="register">ESP</varname> (which is the stack
pointer register), we have effectively
given the microprocessor a minor case
of <emphasis>amnesia</emphasis>: It no longer
remembers it was <function>getchar</function>
that <function role="opcode">call</function>ed <function>read</function>.
</para>

<para>
And since <function>getchar</function> never
<function role="opcode">push</function>ed anything before
<function role="opcode">call</function>ing <function>read</function>,
the top of the stack now contains the
return address to whatever or whoever
<function role="opcode">call</function>ed <function>getchar</function>.
As far as that caller is concerned,
he <function role="opcode">call</function>ed <function>getchar</function>,
which <function role="opcode">ret</function>urned with the
<varname role="register">carry flag</varname> set!
</para>

</blockquote>
<para>
Other than that, the <function>bcdload</function>
routine is caught up in the middle of a
Lilliputian conflict between the Big&ndash;Endians
and the Little&ndash;Endians.
</para>

<para>
It is converting the text representation
of a number into that number: The text
is stored in the big&ndash;endian order, but
the <emphasis>packed decimal</emphasis> is little&ndash;endian.
</para>

<para>
To solve the conflict, we use the <function>std</function>
op code early on. We cancel it with <function>cld</function>
later on: It is quite important we do not
<function>call</function> anything that may depend on
the default setting of the <emphasis>direction
flag</emphasis> while <function>std</function> is active.
</para>

<para>
Everything else in this code should be quite
clear, providing you have read the entire chapter
that precedes it.
</para>

<para>
It is a classical example of the adage that
programming requires a lot of thought and only
a little coding. Once we have thought through every
tiny detail, the code almost writes itself.
</para>

</sect2>

<sect2 id="x86-pinhole-using">
<title>Using <application>pinhole</application></title>
<para>
Because we have decided to make the program
<emphasis>ignore</emphasis> any input except for numbers
(and even those inside a comment), we can
actually perform <emphasis>textual queries</emphasis>.
We do not <emphasis>have to</emphasis>, but we <emphasis>can</emphasis>.
</para>

<para>
In my humble opinion, forming a textual query,
instead of having to follow a very strict
syntax, makes software much more user friendly.
</para>

<para>
Suppose we want to build a pinhole camera to use the
4x5 inch film. The standard focal
length for that film is about 150mm. We want
to <emphasis>fine&ndash;tune</emphasis> our focal length so the
pinhole diameter is as round a number as possible.
Let us also suppose we are quite comfortable with
cameras but somewhat intimidated by computers.
Rather than just have to type in a bunch of numbers,
we want to <emphasis>ask</emphasis> a couple of questions.
</para>

<para>
Our session might look like this:</para>

<screen>&prompt.user; <userinput>pinhole

Computer,

What size pinhole do I need for the focal length of 150?</userinput>
150	490	306	362	2930	12
<userinput>Hmmm... How about 160?</userinput>
160	506	316	362	3125	12
<userinput>Let's make it 155, please.</userinput>
155	498	311	362	3027	12
<userinput>Ah, let's try 157...</userinput>
157	501	313	362	3066	12
<userinput>156?</userinput>
156	500	312	362	3047	12
<userinput>That's it! Perfect! Thank you very much!
^D</userinput></screen>
<para>
We have found that while for the focal length
of 150, our pinhole diameter should be 490
microns, or 0.49 mm, if we go with the almost
identical focal length of 156 mm, we can
get away with a pinhole diameter of exactly
one half of a millimeter.
</para>

</sect2>

<sect2 id="x86-pinhole-scripting">
<title>Scripting</title>
<para>
Because we have chosen the <constant>#</constant>
character to denote the start of a comment,
we can treat our <application>pinhole</application>
software as a <emphasis>scripting language</emphasis>.
</para>

<para>
You have probably seen <application>shell</application>
<emphasis>scripts</emphasis> that start with:</para>

<programlisting>
#! /bin/sh
</programlisting>
<para>
...or...</para>

<programlisting>
#!/bin/sh
</programlisting> <para>
...because the blank space after the <function>#!</function>
is optional.
</para>

<para>
Whenever Unix is asked to run an executable
file which starts with the <function>#!</function>,
it assumes the file is a script. It adds the
command to the rest of the first line of the
script, and tries to execute that.
</para>

<para>
Suppose now that we have installed <application>pinhole</application>
in <application>/usr/local/bin/</application>, we can now
write a script to calculate various pinhole
diameters suitable for various focal lengths
commonly used with the 120 film.</para>

<para>
The script might look something like this:</para>

<programlisting>
#! /usr/local/bin/pinhole -b -i
# Find the best pinhole diameter
# for the 120 film

### Standard
80

### Wide angle
30, 40, 50, 60, 70

### Telephoto
100, 120, 140
</programlisting>
<para>
Because 120 is a medium size film,
we may name this file <application>medium</application>.
</para>

<para>
We can set its permissions to execute,
and run it as if it were a program:
</para>

<screen>&prompt.user; <userinput>chmod 755 medium</userinput>
&prompt.user; <userinput>./medium</userinput></screen>
<para>
Unix will interpret that last command as:</para>

<screen>&prompt.user; <userinput>/usr/local/bin/pinhole -b -i ./medium</userinput></screen>
<para>
It will run that command and display:
</para>

<screen>80	358	224	256	1562	11
30	219	137	128	586	9
40	253	158	181	781	10
50	283	177	181	977	10
60	310	194	181	1172	10
70	335	209	181	1367	10
100	400	250	256	1953	11
120	438	274	256	2344	11
140	473	296	256	2734	11</screen>
<para>

Now, let us enter:</para>

<screen>&prompt.user; <userinput>./medium -c</userinput></screen>
<para>
Unix will treat that as:</para>

<screen>&prompt.user; <userinput>/usr/local/bin/pinhole -b -i ./medium -c</userinput></screen>
<para>
That gives it two conflicting options:
<parameter>-b</parameter> and <parameter>-c</parameter>
(Use Bender's constant and use Connors'
constant). We have programmed it so
later options override early ones&mdash;our
program will calculate everything
using Connors' constant:
</para>

<screen>80	331	242	256	1826	11
30	203	148	128	685	9
40	234	171	181	913	10
50	262	191	181	1141	10
60	287	209	181	1370	10
70	310	226	256	1598	11
100	370	270	256	2283	11
120	405	296	256	2739	11
140	438	320	362	3196	12</screen>
<para>
We decide we want to go with Bender's
constant after all. We want to save its
values as a comma&ndash;separated file:
</para>

<screen>&prompt.user; <userinput>./medium -b -e &gt; bender</userinput>
&prompt.user; <userinput>cat bender</userinput>
focal length in millimeters,pinhole diameter in microns,F-number,normalized F-number,F-5.6 multiplier,stops from F-5.6
80,358,224,256,1562,11
30,219,137,128,586,9
40,253,158,181,781,10
50,283,177,181,977,10
60,310,194,181,1172,10
70,335,209,181,1367,10
100,400,250,256,1953,11
120,438,274,256,2344,11
140,473,296,256,2734,11
&prompt.user;</screen>
</sect2>

</sect1>

<sect1 id="x86-caveats">
<title>Caveats</title>

<para>
Assembly language programmers who "grew up" under
<acronym>MS DOS</acronym> and Windows often tend to take shortcuts.
Reading the keyboard scan codes and writing directly to video
memory are two classical examples of practices which, under
<acronym>MS DOS</acronym> are not frowned upon but considered the
right thing to do.
</para>

<para>
The reason? Both the <acronym>PC BIOS</acronym> and
<acronym>MS DOS</acronym> are notoriously
slow when performing these operations.
</para>

<para>
You may be tempted to continue similar practices in the
Unix environment. For example, I have seen a web site which
explains how to access the keyboard scan codes on a popular Unix clone.
</para>

<para>
That is generally a <emphasis>very bad idea</emphasis>
in Unix environment! Let me explain why.
</para>

<sect2 id="x86-protected">
<title>Unix Is Protected</title>

<para>
For one thing, it may simply not be possible. Unix runs in
protected mode. Only the kernel and device drivers are allowed
to access hardware directly. Perhaps a particular Unix clone
will let you read the keyboard scan codes, but chances are a real
Unix operating system will not. And even if one version may let you
do it, the next one may not, so your carefully crafted software may
become a dinosaur overnight.
</para>

</sect2>

<sect2 id="x86-abstraction">
<title>Unix Is an Abstraction</title>

<para>
But there is a much more important reason not to try
accessing the hardware directly (unless, of course,
you are writing a device driver), even on the Unix-like
systems that let you do it:
</para>

<para>
<emphasis>Unix is an abstraction!
</emphasis></para>

<para>
There is a major difference in the philosophy of design
between <acronym>MS DOS</acronym> and Unix.
<acronym>MS DOS</acronym> was designed as a single-user
system. It is run on a computer with a keyboard and a video
screen attached directly to that computer. User input is almost
guaranteed to come from that keyboard. Your program's output
virtually always ends up on that screen.
</para>

<para>
This is NEVER guaranteed under Unix. It is quite common
for a Unix user to pipe and redirect program input and output:
</para>

<screen>&prompt.user; <userinput>program1 | program2 | program3 > file1</userinput></screen>

<para>
If you have written <application>program2</application>, your input
does not come from the keyboard but from the output of
<application>program1</application>. Similarly, your output does not
go to the screen but becomes the input for
<application>program3</application> whose output, in turn,
goes to <filename>file1</filename>.
</para>

<para>
But there is more! Even if you made sure that your input comes
from, and your output goes to, the terminal, there is no guarantee
the terminal is a PC: It may not have its video memory
where you expect it, nor may its keyboard be producing
<acronym>PC</acronym>-style scan codes. It may be a Macintosh,
or any other computer.
</para>

<para>
Now you may be shaking your head: My software is in
<acronym>PC</acronym> assembly language, how can
it run on a Macintosh? But I did not say your software
would be running on a Macintosh, only that its terminal
may be a Macintosh.
</para>

<para>
Under Unix, the terminal does not have to be directly
attached to the computer that runs your software, it can
even be on another continent, or, for that matter, on another
planet. It is perfectly possible that a Macintosh user in
Australia connects to a Unix system in North America (or
anywhere else) via <application>telnet</application>. The
software then runs on one computer, while the terminal is
on a different computer: If you try to read the scan codes,
you will get the wrong input!
</para>

<para>
Same holds true about any other hardware: A file you are reading
may be on a disk you have no direct access to. A camera you are
reading images from may be on a space shuttle, connected to you
via satellites.
</para>

<para>
That is why under Unix you must never make any assumptions about
where your data is coming from and going to. Always let the
system handle the physical access to the hardware.
</para>

<note>
<para>
These are caveats, not absolute rules. Exceptions are possible.
For example, if a text editor has determined it is running
on a local machine, it may want to read the scan codes
directly for improved control. I am not mentioning these caveats
to tell you what to do or what not to do, just to make you aware
of certain pitfalls that await you if you have just arrived to Unix
form <acronym>MS DOS</acronym>. Of course, creative people often break
rules, and it is OK as long as they know they are breaking
them and why.
</para>
</note>

</sect2>

</sect1>


<sect1 id="x86-acknowledgements">
<title>Acknowledgements</title>

<para>
This tutorial would never have been possible without the
help of many experienced FreeBSD programmers from the
<ulink url="mailto:freebsd-hackers@FreeBSD.org">FreeBSD
hackers</ulink> mailing list, many of whom have patiently
answered my questions, and pointed me in the right direction
in my attempts to explore the inner workings of Unix
system programming in general and FreeBSD in particular.
</para>

<para>
Thomas M. Sommers opened the door for me. His
<ulink url="http://home.ptd.net/~tms2/hello.html">How
do I write "Hello, world" in FreeBSD assembler?</ulink>
web page was my first encounter with an example of
assembly language programming under FreeBSD.
</para>

<para>
Jake Burkholder has kept the door open by willingly
answering all of my questions and supplying me with
example assembly language source code.
</para>

<para>
Copyright &copy; 2000-2001 G. Adam Stanislav. All rights reserved.
</para>

</sect1>


</chapter>