Art of Assembly/Win32 Edition is now available. Let me read that version.
PLEASE: Before emailing me asking how to get a hard copy of this text, read this.
Important Notice: As you have probably discovered by now, I am no longer updating this document. The reason is quite simple: I'm working on a Windows version of "The Art of Assembly Language Programming". In the past I have encouraged individuals to send me corrections to this text. However, as I am no longer updating this material. However, I am collecting errata that I will post to Webster someday, so feel free to continue sending corrections to AoA/DOS (16-bit) to rhyde@cs.ucr.edu. If you're more interested in leading edge material, please see the information about the Win/32 edition, above.
mov) that will prove handy in the discussion throughout the rest of this chapter. These are the load effective address (lea), load es and general purpose register (les), addition (add), and multiply (mul). These instructions, along with the mov instruction, provide all the necessary power to access the different data types this chapter discusses.lea instruction takes the form:
lea reg16, memory
reg16 is a 16 bit general purpose register. Memory is a memory location represented by a mod/reg/rm byte (except it must be a memory location, it cannot be a register).
This instruction loads the 16 bit register with the offset of the location specified by the memory operand. lea ax,1000h[bx][si], for example, would load ax with the address of the memory location pointed at by 1000h[bx][si]. This, of course, is the value 1000h+bx+si. Lea is also quite useful for obtaining the address of a variable. If you have a variable I somewhere in memory, lea bx,I will load the bx register with the address (offset) of I.
The les instruction takes the form
les reg16, memory32
This instruction loads the es register and one of the 16 bit general purpose registers from the specified memory address. Note that any memory address you can specify with a mod/reg/rm byte is legal but like the lea instruction it must be a memory location, not a register.
The les instruction loads the specified general purpose register from the word at the given address, it loads the es register from the following word in memory. This instruction, and it's companion lds (which loads ds) are the only instructions on pre-80386 machines that manipulate 32 bits at a time.
The add instruction, like it's x86 counterpart, adds two values on the 80x86. This instruction takes several forms. There are five forms that concern us here. They are
add reg, reg
add reg, memory
add memory, reg
add reg, constant
add memory, constant
All these instructions add the second operand to the first leaving the sum in the first operand. For example, add bx,5 computes bx := bx + 5.
The last instruction to look at is the mul (multiply) instruction. This instruction has only a single operand and takes the form:
mul reg/memory
There are many important details concerning mul that this chapter ignores. For the sake of the discussion that follows, assume that the register or memory location is a 16 bit register or memory location. In such a case this instruction computes dx:ax :=ax*reg/mem. Note that there is no immediate mode for this instruction.
program useless(input,output);
var i,j:integer;
begin
i := 10;
write('Enter a value for j:');
readln(j);
i := i*j + j*j;
writeln('The result is ',i);
end.
When the computer executes the statement i:=10; it makes a copy of the value 10 and somehow remembers this value for use later on. To accomplish this, the compiler sets aside a memory location specifically for the exclusive use of the variable i. Assuming the compiler arbitrarily assigned location DS:10h for this purpose it could use the instruction mov ds:[10h],10 to accomplish this. If i is a 16 bit word, the compiler would probably assign the variable j to the word starting at location 12h or 0Eh. Assuming it's location 12h, the second assignment statement in the program might wind up looking like the following:
mov ax, ds:[10h] ;Fetch value of I
mul ds:[12h] ;Multiply by J
mov ds:[10h], ax ;Save in I (ignore overflow)
mov ax, ds:[12h] ;Fetch J
mul ds:[12h] ;Compute J*J
add ds:[10h], ax ;Add I*J + J*J, store into I
Although there are a few details missing from this code, it is fairly straightforward and you can easily see what is going on in this program.
Now imagine a 5,000 line program like this one using variables like ds:[10h], ds:[12h], ds:[14h], etc. Would you want to locate the statement where you accidentally stored the result of a computation into j rather than i? Indeed, why should you even care that the variable i is at location 10h and j is at location 12h? Why shouldn't you be able to use names like i and j rather than worrying about these numerical addresses? It seems reasonable to rewrite the code above as:
mov ax, i
mul j
mov i, ax
mov ax, j
mul j
add i, ax
Of course you can do this in assembly language! Indeed, one of the primary jobs of an assembler like MASM is to let you use symbolic names for memory locations. Furthermore, the assembler will even assign locations to the names automatically for you. You needn't concern yourself with the fact that variable i is really the word at memory location DS:10h unless you're curious.
It should come as no surprise that ds will point to the dseg segment in the SHELL.ASM file. Indeed, setting up ds so that it points at dseg is one of the first things that happens in the SHELL.ASM main program. Therefore, all you've got to do is tell the assembler to reserve some storage for your variables in dseg and attach the offset of said variables with the names of those variables. This is a very simple process and is the subject of the next several sections.
i and j in the preceding section are examples of scalar variables. Examples of data structures that are not scalars include arrays, records, sets, and lists. These latter data types are made up from scalar values. They are the composite types. You'll see the composite types a little later; first you need to learn to deal with the scalar types.ByteVar byte ?
ByteVar is a label. It should begin at column one on the line somewhere in the dseg segment (that is, between the dseg segment and dseg ends statements). You'll find out all about labels in a few chapters, for now you can assume that most legal Pascal/C/Ada identifiers are also valid assembly language labels.
If you need more than one variable in your program, just place additional lines in the dseg segment declaring those variables. MASM will automatically allocate a unique storage location for the variable (it wouldn't be too good to have i and j located at the same address now, would it?). After declaring said variable, MASM will allow you to refer to that variable by name rather than by location in your program. For example, after inserting the above statement into the data segment (dseg), you could use instructions like mov ByteVar,al in your program.
The first variable you place in the data segment gets allocated storage at location DS:0. The next variable in memory gets allocated storage just beyond the previous variable. For example, if the variable at location zero was a byte variable, the next variable gets allocated storage at DS:1. However, if the first variable was a word, the second variable gets allocated storage at location DS:2. MASM is always careful to allocate variables in such a manner that they do not overlap. Consider the following dseg definition:
dseg segment para public 'data' bytevar byte ? ;byte allocates bytes wordvar word ? ;word allocates words dwordvar dword ? ;dword allocs dbl words byte2 byte ? word2 word ? dseg ends
MASM allocates storage for bytevar at location DS:0. Because bytevar is one byte long, the next available memory location is going to be DS:1. MASM, therefore, allocates storage for wordvar at location DS:1. Since words require two bytes, the next available memory location after wordvar is DS:3 which is where MASM allocates storage for dwordvar. Dwordvar is four bytes long, so MASM allocates storage for byte2 starting at location DS:7. Likewise, MASM allocates storage for word2 at location DS:8. Were you to stick another variable after word2, MASM would allocate storage for it at location DS:0A.
Whenever you refer to one of the names above, MASM automatically substitutes the appropriate offset. For example, MASM would translate the mov ax,wordvar instruction into mov ax,ds:[1]. So now you can use symbolic names for your variables and completely ignore the fact that these variables are really memory locations with corresponding offsets into the data segment.
colors = (red, blue, green, purple, orange, yellow, white, black);
Most Pascal compilers will assign the value zero to red, one to blue, two to green, etc.
Later, you will see how to actually create your own enumerated data types in assembly language. All you need to learn now is how to allocate storage for a variable that holds an enumerated value. Since it's unlikely there will be more than 256 items enumerated by the data type, you can use a simple byte variable to hold the value. If you have a variable, say color of type colors, using the instruction mov color,2 is the same thing as saying color:=green in Pascal. (Later, you'll even learn how to use more meaningful statements like mov color,green to assign the color green to the color variable).
Of course, if you have a small unsigned integer value (0...255) or small signed integer (-128...127) a single byte variable is the best way to go in most cases. Note that most programmers treat all data types except small signed integers as unsigned values. That is, characters, booleans, enumerated types, and unsigned integers are all usually unsigned values. In some very special cases you might want to treat a character as a signed value, but most of the time even characters are unsigned values.
There are three main statements for declaring byte variables in a program. They are
identifier db ? identifier byte ? and identifier sbyte ?
identifier represents the name of your byte variable. "db" is an older term that predates MASM 6.x. You will see this directive used quite a bit by other programmers (especially those who are not using MASM 6.x or later) but Microsoft considers it to be an obsolete term; you should always use the byte and sbyte declarations instead.
The byte declaration declares unsigned byte variables. You should use this declaration for all byte variables except small signed integers. For signed integer values, use the sbyte (signed byte) directive.
Once you declare some byte variables with these statements, you may reference those variables within your program by their names:
i db ?
j byte ?
k sbyte ?
.
.
.
mov i, 0
mov j, 245
mov k, -5
mov al, i
mov j, al
etc.
Although MASM 6.x performs a small amount of type checking, you should not get the idea that assembly language is a strongly typed language. In fact, MASM 6.x will only check the values you're moving around to verify that they will fit in the target location. All of the following are legal in MASM 6.x:
mov k, 255
mov j, -5
mov i, -127
Since all of these variables are byte-sized variables, and all the associated constants will fit into eight bits, MASM happily allows each of these statements. Yet if you look at them, they are logically incorrect. What does it mean to move -5 into an unsigned byte variable? Since signed byte values must be in the range -128...127, what happens when you store the value 255 into a signed byte variable? Well, MASM simply converts these values to their eight bit equivalents (-5 becomes 0FBh, 255 becomes 0FFh [-1], etc.).
Perhaps a later version of MASM will perform stronger type checking on the values you shove into these variables, perhaps not. However, you should always keep in mind that it will always be possible to circumvent this checking. It's up to you to write your programs correctly. The assembler won't help you as much as Pascal or Ada will. Of course, even if the assembler disallowed these statements, it would still be easy to get around the type checking. Consider the following sequence:
mov al, -5
.
; Any number of statements which do not affect AL
.
mov j, al
There is, unfortunately, no way the assembler is going to be able to tell you that you're storing an illegal value into j. The registers, by their very nature, are neither signed nor unsigned. Therefore the assembler will let you store a register into a variable regardless of the value that may be in that register.
Although the assembler does not check to see if both operands to an instruction are signed or unsigned, it most certainly checks their size. If the sizes do not agree the assembler will complain with an appropriate error message. The following examples are all illegal:
mov i, ax ;Cannot move 16 bits into eight
mov i, 300 ;300 won't fit in eight bits.
mov k, -130 ;-130 won't fit into eight bits.
You might ask "if the assembler doesn't really differentiate signed and unsigned values, why bother with them? Why not simply use db all the time?" Well, there are two reasons. First, it makes your programs easier to read and understand if you explicitly state (by using byte and sbyte) which variables are signed and which are unsigned. Second, who said anything about the assembler ignoring whether the variables are signed or unsigned? The mov instruction ignores the difference, but there are other instructions that do not.
One final point is worth mentioning concerning the declaration of byte variables. In all of the declarations you've seen thus far the operand field of the instruction has always contained a question mark. This question mark tells the assembler that the variable should be left uninitialized when DOS loads the program into memory. You may specify an initial value for the variable, that will be loaded into memory before the program starts executing, by replacing the question mark with your initial value. Consider the following byte variable declarations:
i db 0 j byte 255 k sbyte -1
In this example, the assembler will initialize i, j, and k to zero, 255, and -1, respectively, when the program loads into memory. This fact will prove quite useful later on, especially when discussing tables and arrays. Once again, the assembler only checks the sizes of the operands. It does not check to make sure that the operand for the byte directive is positive or that the value in the operand field of sbyte is in the range -128...127. MASM will allow any value in the range -128...255 in the operand field of any of these statements.
In case you get the impression that there isn't a real reason to use byte vs. sbyte in a program, you should note that while MASM sometimes ignores the differences in these definitions, Microsoft's CodeView debugger does not. If you've declared a variable as a signed value, CodeView will display it as such (including a minus sign, if necessary). On the other hand, CodeView will always display db and byte variables as positive values.
dw, word, and sword statements to declare word variables. The following examples demonstrate their use:
NoSignedWord dw ? UnsignedWord word ? SignedWord sword ? Initialized0 word 0 InitializedM1 sword -1 InitializedBig word 65535 InitializedOfs dw NoSignedWord
Most of these declarations are slight modifications of the byte declarations you saw in the last section. Of course you may initialize any word variable to a value in the range -32768...65535 (the union of the range for signed and unsigned 16 bit constants). The last declaration above, however, is new. In this case a label appears in the operand field (specifically, the name of the NoSignedWord variable). When a label appears in the operand field the assembler will substitute the offset of that label (within the variable's segment). If these were the only declarations in dseg and they appeared in this order, the last declaration above would initialize InitializedOfs with the value zero since NoSignedWord's offset is zero within the data segment. This form of initialization is quite useful for initializing pointers. But more on that subject later.
The CodeView debugger differentiates dw/word variables and sword variables. It always displays the unsigned values as positive integers. On the other hand, it will display sword variables as signed values (complete with minus sign, if the value is negative). Debugging support is one of the main reasons you'll want to use word or sword as appropriate.
dd, dword, and sdword instructions to declare four-byte integers, pointers, and other variables types. Such variables will allow values in the range -2,147,483,648...4,294,967,295 (the union of the range of signed and unsigned four-byte integers). You use these declarations like the word declarations:
NoSignedDWord dd ? UnsignedDWord dword ? SignedDWord sdword ? InitBig dword 4000000000 InitNegative sdword -1 InitPtr dd InitBig
The last example initializes a double word pointer with the segment:offset address of the InitBig variable.
Once again, it's worth pointing out that the assembler doesn't check the types of these variables when looking at the initialization values. If the value fits into 32 bits, the assembler will accept it. Size checking, however, is strictly enforced. Since the only 32 bit mov instructions on processors earlier than the 80386 are les and lds, you will get an error if you attempt to access dword variables on these earlier processors using a mov instruction. Of course, even on the 80386 you cannot move a 32 bit variable into a 16 bit register, you must use the 32 bit registers. Later, you'll learn how to manipulate 32 bit variables, even on a 16 bit processor. Until then, just pretend that you can't.
Keep in mind, of course, that CodeView differentiates between dd/dword and sdword. This will help you see the actual values your variables have when you're debugging your programs. CodeView only does this, though, if you use the proper declarations for your variables. Always use sdword for signed values and dd or dword (dword is best) for unsigned values.
df/fword, dq/qword, and dt/tbyte statements. Declarations using these statements were originally intended for floating point and BCD values. There are better directives for the floating point variables and you don't need to concern yourself with the other data types you'd use these directives for. The following discussion is for completeness' sake.df/fword statement's main utility is declaring 48 bit pointers for use in 32 bit protected mode on the 80386 and later. Although you could use this directive to create an arbitrary six byte variable, there are better directives for doing that. You should only use this directive for 48 bit far pointers on the 80386.dq/qword lets you declare quadword (eight byte) variables. The original purpose of this directive was to let you create 64 bit double precision floating point variables and 64 bit integer variables. There are better directives for creating floating point variables. As for 64 bit integers, you won't need them very often on the 80x86 CPU (at least, not until Intel releases a member of the 80x86 family with 64 bit general purpose registers).dt/tbyte directives allocate ten bytes of storage. There are two data types indigenous to the 80x87 (math coprocessor) family that use a ten byte data type: ten byte BCD values and extended precision (80 bit) floating point values. This text will pretty much ignore the BCD data type. As for the floating point type, once again there is a better way to do it.
dd, dq, and dt these statements reserve four, eight, and ten bytes. The operand fields for these statements may contain a question mark (if you don't want to initialize the variable) or it may contain an initial value in floating point form. The following examples demonstrate their use:
x real4 1.5 y real8 1.0e-25 z real10 -1.2594e+10
Note that the operand field must contain a valid floating point constant using either decimal or scientific notation. In particular, pure integer constants are not allowed. The assembler will complain if you use an operand like the following:
x real4 1
To correct this, change the operand field to "1.0".
Please note that it takes special hardware to perform floating point operations (e.g., an 80x87 chip or an 80x86 with built-in math coprocessor). If such hardware is not available, you must write software to perform operations like floating point addition, subtraction, multiplication, etc. In particular, you cannot use the 80x86 add instruction to add two floating point values. This text will cover floating point arithmetic in a later chapter. Nonetheless, it's appropriate to discuss how to declare floating point variables in the chapter on data structures.
MASM also lets you use dd, dq, and dt to declare floating point variables (since these directives reserve the necessary four, eight, or ten bytes of space). You can even initialize such variables with floating point constants in the operand field. But there are two major drawbacks to declaring variables this way. First, as with bytes, words, and double words, the CodeView debugger will only display your floating point variables properly if you use the real4, real8, or real10 directives. If you use dd, dq, or dt, CodeView will display your values as four, eight, or ten byte unsigned integers. Another, potentially bigger, problem with using dd, dq, and dt is that they allow both integer and floating point constant initializers (remember, real4, real8, and real10 do not). Now this might seem like a good feature at first glance. However, the integer representation for the value one is not the same as the floating point representation for the value 1.0. So if you accidentally enter the value "1" in the operand field when you really meant "1.0", the assembler would happily digest this and then give you incorrect results. Hence, you should always use the real4, real8, and real10 statements to declare floating point variables.
integer typedef sword char typedef byte boolean typedef byte float typedef real4 colors typedef byte
Now you can declare your variables with more meaningful statements like:
i integer ? ch char ? FoundIt boolean ? x float ? HouseColor colors ?
If you are an Ada, C, or FORTRAN programmer (or any other language, for that matter), you can pick type names you're more comfortable with. Of course, this doesn't change how the 80x86 or MASM reacts to these variables one iota, but it does let you create programs that are easier to read and understand since the type names are more indicative of the actual underlying types.
Note that CodeView still respects the underlying data type. If you define integer to be an sword type, CodeView will display variables of type integer as signed values. Likewise, if you define float to mean real4, CodeView will still properly display float variables as four-byte floating point values.
M: array [0..1023] of integer;
Even if you don't know Pascal, the concept here is pretty easy to understand. M is an array with 1024 integers in it, indexed from M[0] to M[1023]. Each one of these array elements can hold an integer value that is independent of all the others. In other words, this array gives you 1024 different integer variables each of which you refer to by number (the array index) rather than by name.
If you encountered a program that had the statement M[0]:=100 you probably wouldn't have to think at all about what is happening with this statement. It is storing the value 100 into the first element of the array M. Now consider the following two statements:
i := 0; (* Assume "i" is an integer variable *) M [i] := 100;
You should agree, without too much hesitation, that these two statements perform the same exact operation as M[0]:=100;. Indeed, you're probably willing to agree that you can use any integer expression in the range 0...1023 as an index into this array. The following statements still perform the same operation as our single assignment to index zero:
i := 5; (* assume all variables are integers*)
j := 10;
k := 50;
m [i*j-k] := 100;
"Okay, so what's the point?" you're probably thinking. "Anything that produces an integer in the range 0...1023 is legal. So what?" Okay, how about the following:
M [1] := 0; M [ M [1] ] := 100;
Whoa! Now that takes a few moments to digest. However, if you take it slowly, it makes sense and you'll discover that these two instructions perform the exact same operation you've been doing all along. The first statement stores zero into array element M[1]. The second statement fetches the value of M[1], which is an integer so you can use it as an array index into M, and uses that value (zero) to control where it stores the value 100.
If you're willing to accept the above as reasonable, perhaps bizarre, but usable nonetheless, then you'll have no problems with pointers. Because m[1] is a pointer! Well, not really, but if you were to change "M" to "memory" and treat this array as all of memory, this is the exact definition of a pointer.
A pointer is simply a memory location whose value is the address (or index, if you prefer) of some other memory location. Pointers are very easy to declare and use in an assembly language program. You don't even have to worry about array indices or anything like that. In fact, the only complication you're going to run into is that the 80x86 supports two kinds of pointers: near pointers and far pointers.
A near pointer is a 16 bit value that provides an offset into a segment. It could be any segment but you will generally use the data segment (dseg in SHELL.ASM). If you have a word variable p that contains 1000h, then p "points" at memory location 1000h in dseg. To access the word that p points at, you could use code like the following:
mov bx, p ;Load BX with pointer.
mov ax, [bx] ;Fetch data that p points at.
By loading the value of p into bx this code loads the value 1000h into bx (assuming p contains 1000h and, therefore, points at memory location 1000h in dseg). The second instruction above loads the ax register with the word starting at the location whose offset appears in bx. Since bx now contains 1000h, this will load ax from locations DS:1000 and DS:1001.
Why not just load ax directly from location 1000h using an instruction like mov ax,ds:[1000h]? Well, there are lots of reasons. But the primary reason is that this single instruction always loads ax from location 1000h. Unless you are willing to mess around with self-modifying code, you cannot change the location from which it loads ax. The previous two instructions, however, always load ax from the location that p points at. This is very easy to change under program control, without using self-modifying code. In fact, the simple instruction mov p,2000h will cause those two instructions above to load ax from memory location DS:2000 the next time they execute. Consider the following instructions:
lea bx, i ;This can actually be done with
mov p, bx ; a single instruction as you'll
. ; see in Chapter Eight.
.
< Some code that skips over the next two instructions >
lea bx, j ;Assume the above code skips these
mov p, bx ; two instructions, that you get
. ; here by jumping to this point from
. ; somewhere else.
mov bx, p ;Assume both code paths above wind
mov ax, [bx] ; up down here.
This short example demonstrates two execution paths through the program. The first path loads the variable p with the address of the variable i (remember, lea loads bx with the offset of the second operand). The second path through the code loads p with the address of the variable j. Both execution paths converge on the last two mov instructions that load ax with i or j depending upon which execution path was taken. In many respects, this is like a parameter to a procedure in a high level language like Pascal. Executing the same instructions accesses different variables depending on whose address (i or j) winds up in p.
Sixteen bit near pointers are small, fast, and the 80x86 provides efficient access using them. Unfortunately, they have one very serious drawback - you can only access 64K of data (one segment) when using near pointers. Far pointers overcome this limitation at the expense of being 32 bits long. However, far pointers let you access any piece of data anywhere in the memory space. For this reason, and the fact that the UCR Standard Library uses far pointers exclusively, this text will use far pointers most of the time. But keep in mind that this is a decision based on trying to keep things simple. Code that uses near pointers rather than far pointers will be shorter and faster.
To access data referenced by a 32 bit pointer, you will need to load the offset portion (L.O. word) of the pointer into bx, bp, si, or di and the segment portion into a segment register (typically es). Then you could access the object using the register indirect addressing mode. Since the les instruction is so convenient for this operation, it is the perfect choice for loading es and one of the above four registers with a pointer value. The following sample code stores the value in al into the byte pointed at by the far pointer p:
les bx, p ;Load p into ES:BX
mov es:[bx], al ;Store away AL
Since near pointers are 16 bits long and far pointers are 32 bits long, you could simply use the dw/word and dd/dword directives to allocate storage for your pointers (pointers are inherently unsigned, so you wouldn't normally use sword or sdword to declare a pointer). However, there is a much better way to do this by using the typedef statement. Consider the following general forms:
typename typedef near ptr basetype typename typedef far ptr basetype
In these two examples typename represents the name of the new type you're creating while basetype is the name of the type you want to create a pointer for. Let's look at some specific examples:
nbytptr typedef near ptr byte fbytptr typedef far ptr byte colorsptr typedef far ptr colors wptr typedef near ptr word intptr typedef near ptr integer intHandle typedef near ptr intptr
(these declarations assume that you've previously defined the types colors and integer with the typedef statement). The typedef statements with the near ptr operands produce 16 bit near pointers. Those with the far ptr operands produce 32 bit far pointers. MASM 6.x ignores the base type supplied after the near ptr or far ptr. However, CodeView uses the base type to display the object a pointer refers to in its correct format.
Note that you can use any type as the base type for a pointer. As the last example above demonstrates, you can even define a pointer to another pointer (a handle). CodeView would properly display the object a variable of type intHandle points at as an address.
With the above types, you can now generate pointer variables as follows:
bytestr nbytptr ? bytestr2 fbytptr ? CurrentColor colorsptr ? CurrentItem wptr ? LastInt intptr ?
Of course, you can initialize these pointers at assembly time if you know where they are going to point when the program first starts running. For example, you could initialize the bytestr variable above with the offset of MyString using the following declaration:
bytestr nbytptr MyString