10.3 Arithmetic Expressions
Probably the biggest shock to beginners facing assembly language for the very first time is the lack of familiar arithmetic expressions. Arithmetic expressions, in most high level languages, look similar to their algebraic equivalents, e.g.,
X:=Y*Z;In assembly language, you'll need several statements to accomplish this same task, e.g.,
mov( y, eax ); intmul( z, eax ); mov( eax, x );Obviously the HLL version is much easier to type, read, and understand. This point, more than any other, is responsible for scaring people away from assembly language.
Although there is a lot of typing involved, converting an arithmetic expression into assembly language isn't difficult at all. By attacking the problem in steps, the same way you would solve the problem by hand, you can easily break down any arithmetic expression into an equivalent sequence of assembly language statements. By learning how to convert such expressions to assembly language in three steps, you'll discover there is little difficulty to this task.
10.3.1 Simple Assignments
The easiest expressions to convert to assembly language are the simple assignments. Simple assignments copy a single value into a variable and take one of two forms:
variable := constantvariable := variableConverting the first form to assembly language is trivial, just use the assembly language statement:
mov( constant, variable );This MOV instruction copies the constant into the variable.
The second assignment above is slightelly more complicated since the 80x86 doesn't provide a memorytomemory MOV instruction. Therefore, to copy one memory variable into another, you must move the data through a register. By convention (and for slight efficiency reasons), most programmers tend to use AL/AX/EAX for this purpose. If the AL, AX, or EAX register is available, you should use it for this operation. For example,
var1 := var2;mov( var2, eax ); mov( eax, var1 );This is assuming, of course, that var1 and var2 are 32bit variables. Use AL if they are eight bit variables, use AX if they are 16bit variables.
Of course, if you're already using AL, AX, or EAX for something else, one of the other registers will suffice. Regardless, you must use a register to transfer one memory location to another.
Although the 80x86 does not support a memorytomemory move, HLA does provide an extended syntax for the MOV instruction that allows two memory operands. However, both operands have to be 16bit or 32bit values; eightbit values won't work. Assuming you want to copy the value of a word or dword object to another variable, you can use the following syntax:
mov( var2, var1 );HLA translates this "instruction" into the following two instruction sequence:
push( var2 ); pop( var1 );Although this is slightly slower than the two MOV instructions, it is convenient.
10.3.2 Simple Expressions
The next level of complexity up from a simple assignment is a simple expression. A simple expression takes the form:
var1 := term1 op term2;Var1 is a variable, term1 and term2 are variables or constants, and op is some arithmetic operator (addition, subtraction, multiplication, etc.).
As simple as this expression appears, most expressions take this form. It should come as no surprise then, that the 80x86 architecture was optimized for just this type of expression.
A typical conversion for this type of expression takes the following form:
mov( term1, eax ); op( term2, eax ); mov( eax, var1 )Op is the mnemonic that corresponds to the specified operation (e.g., "+" = add, "" = sub, etc.).
There are a few inconsistencies you need to be aware of. Of course, when dealing with the multiply and divide instructions on the 80x86, you must use the AL/AX/EAX and DX/EDX registers. You cannot use arbitrary registers as you can with other operations. Also, don't forget the sign extension instructions if you're performing a division operation and you're dividing one 16/32 bit number by another. Finally, don't forget that some instructions may cause overflow. You may want to check for an overflow (or underflow) condition after an arithmetic operation.
Examples of common simple expressions:
x := y + z;mov( y, eax ); add( z, eax ); mov( eax, x ); x := y  z;mov( y, eax ); sub( z, eax ); mov( eax, x );x := y * z; {unsigned}mov( y, eax ); mul( z, eax ); // Don't forget this wipes out EDX. mov( eax, x );x := y div z; {unsigned div}mov( y, eax ); mov( 0, edx ); // Zero extend EAX into EDX. div( z, edx:eax ); mov( eax, x );x := y idiv z; {signed div}mov( y, eax ); cdq(); // Sign extend EAX into EDX. idiv( z, edx:eax ); mov( eax, z );x := y mod z; {unsigned remainder}mov( y, eax ); mov( 0, edx ); // Zero extend EAX into EDX. mod( z, edx:eax ); mov( edx, x ); // Note that remainder is in EDX.x := y imod z; {signed remainder}mov( y, eax ); cdq(); // Sign extend EAX into EDX. imod( z, edx:eax ); mov( edx, x ); // Remainder is in EDX.Certain unary operations also qualify as simple expressions. A good example of a unary operation is negation. In a high level language negation takes one of two possible forms:
var := var or var1 := var2Note that var := constant is really a simple assignment, not a simple expression. You can specify a negative constant as an operand to the MOV instruction:
mov( 14, var );To handle "var = var;" use the single assembly language statement:
// var = var; neg( var );If two different variables are involved, then use the following:
// var1 = var2; mov( var2, eax ); neg( eax ); mov( eax, var1 );10.3.3 Complex Expressions
A complex expression is any arithmetic expression involving more than two terms and one operator. Such expressions are commonly found in programs written in a high level language. Complex expressions may include parentheses to override operator precedence, function calls, array accesses, etc. While the conversion of some complex expressions to assembly language is fairly straightforward, others require some effort. This section outlines the rules you use to convert such expressions.
A complex expression that is easy to convert to assembly language is one that involves three terms and two operators, for example:
w := w  y  z;Clearly the straightforward assembly language conversion of this statement will require two SUB instructions. However, even with an expression as simple as this one, the conversion is not trivial. There are actually two ways to convert this from the statement above into assembly language:
mov( w, eax ); sub( y, eax ); sub( z, eax ); mov( eax, w ); and mov( y, eax ); sub( z, eax ); sub( eax, w );The second conversion, since it is shorter, looks better. However, it produces an incorrect result (assuming Pascallike semantics for the original statement). Associativity is the problem. The second sequence above computes W := W  (Y  Z) which is not the same as W := (W  Y)  Z. How we place the parentheses around the subexpressions can affect the result. Note that if you are interested in a shorter form, you can use the following sequence:
mov( y, eax ); add( z, eax ); sub( eax, w );This computes W:=W(Y+Z). This is equivalent to W := (W  Y)  Z.
Precedence is another issue. Consider the Pascal expression:
X := W * Y + Z;Once again there are two ways we can evaluate this expression:
X := (W * Y) + Z; or X := W * (Y + Z);By now, you're probably thinking that this text is crazy. Everyone knows the correct way to evaluate these expressions is the second form provided in these two examples. However, you're wrong to think that way. The APL programming language, for example, evaluates expressions solely from right to left and does not give one operator precedence over another.
Most high level languages use a fixed set of precedence rules to describe the order of evaluation in an expression involving two or more different operators. Such programming languages usually compute multiplication and division before addition and subtraction. Those that support exponentiation (e.g., FORTRAN and BASIC) usually compute that before multiplication and division. These rules are intuitive since almost everyone learns them before high school. Consider the expression:
X op1 Y op2 ZIf op1 takes precedence over op2 then this evaluates to (X op1 Y) op2 Z otherwise if op2 takes precedence over op1 then this evaluates to X op1 (Y op2 Z ). Depending upon the operators and operands involved, these two computations could produce different results.
When converting an expression of this form into assembly language, you must be sure to compute the subexpression with the highest precedence first. The following example demonstrates this technique:
// w := x + y * z;mov( x, ebx ); mov( y, eax ); // Must compute y*z first since "*" intmul( z, eax ); // has higher precedence than "+". add( ebx, eax ); mov( eax, w );If two operators appearing within an expression have the same precedence, then you determine the order of evaluation using associativity rules. Most operators are left associative meaning that they evaluate from left to right. Addition, subtraction, multiplication, and division are all left associative. A right associative operator evaluates from right to left. The exponentiation operator in FORTRAN and BASIC is a good example of a right associative operator:
2^2^3 is equal to 2^(2^3) not (2^2)^3
The precedence and associativity rules determine the order of evaluation. Indirectly, these rules tell you where to place parentheses in an expression to determine the order of evaluation. Of course, you can always use parentheses to override the default precedence and associativity. However, the ultimate point is that your assembly code must complete certain operations before others to correctly compute the value of a given expression. The following examples demonstrate this principle:
// w := x  y  zmov( x, eax ); // All the same operator, so we need sub( y, eax ); // to evaluate from left to right sub( z, eax ); // because they all have the same mov( eax, w ); // precedence and are left associative.// w := x + y * zmov( y, eax ); // Must compute Y * Z first since intmul( z, eax ); // multiplication has a higher add( x, eax ); // precedence than addition. mov( eax, w );// w := x / y  zmov( x, eax ); // Here we need to compute division cdq(); // first since it has the highest idiv( y, edx:eax ); // precedence. sub( z, eax ); mov( eax, w );// w := x * y * zmov( y, eax ); // Addition and multiplication are intmul( z, eax ); // commutative, therefore the order intmul( x, eax ); // of evaluation does not matter mov( eax, w );
There is one exception to the associativity rule. If an expression involves multiplication and division it is generally better to perform the multiplication first. For example, given an expression of the form:
W := X/Y * Z // Note: this is not !It is usually better to compute X*Z and then divide the result by Y rather than divide X by Y and multiply the quotient by Z. There are two reasons this approach is better. First, remember that the IMUL instruction always produces a 64 bit result (assuming 32 bit operands). By doing the multiplication first, you automatically sign extend the product into the EDX register so you do not have to sign extend EAX prior to the division. This saves the execution of the CDQ instruction. A second reason for doing the multiplication first is to increase the accuracy of the computation. Remember, (integer) division often produces an inexact result. For example, if you compute 5/2 you will get the value two, not 2.5. Computing (5/2)*3 produces six. However, if you compute (5*3)/2 you get the value seven which is a little closer to the real quotient (7.5). Therefore, if you encounter an expression of the form:
w := x/y*z;You can usually convert it to the assembly code:
mov( x, eax ); imul( z, eax ); // Note the use of IMUL, not INTMUL! idiv( y, edx:eax ); mov( eax, w );Of course, if the algorithm you're encoding depends on the truncation effect of the division operation, you cannot use this trick to improve the algorithm. Moral of the story: always make sure you fully understand any expression you are converting to assembly language. Obviously if the semantics dictate that you must perform the division first, do so.
Consider the following Pascal statement:
w := x  y * x;This is similar to a previous example except it uses subtraction rather than addition. Since subtraction is not commutative, you cannot compute y * z and then subtract x from this result. This tends to complicate the conversion a tiny amount. Rather than a straight forward multiply and addition sequence, you'll have to load x into a register, multiply y and z leaving their product in a different register, and then subtract this product from x, e.g.,
mov( x, ebx ); mov( y, eax ); intmul( x, eax ); sub( eax, ebx ); mov( ebx, w );This is a trivial example that demonstrates the need for temporary variables in an expression. This code uses the EBX register to temporarily hold a copy of x until it computes the product of y and z. As your expressions increase in complexity, the need for temporaries grows. Consider the following Pascal statement:
w := (a + b) * (y + z);Following the normal rules of algebraic evaluation, you compute the subexpressions inside the parentheses (i.e., the two subexpressions with the highest precedence) first and set their values aside. When you computed the values for both subexpressions you can compute their sum. One way to deal with complex expressions like this one is to reduce it to a sequence of simple expressions whose results wind up in temporary variables. For example, we can convert the single expression above into the following sequence:
Temp1 := a + b; Temp2 := y + z; w := Temp1 * Temp2;Since converting simple expressions to assembly language is quite easy, it's now a snap to compute the former, complex, expression in assembly. The code is
mov( a, eax ); add( b, eax ); mov( eax, Temp1 ); mov( y, eax ); add( z, eax ); mov( eax, Temp2 ); mov( Temp1, eax ); intmul( Temp2, eax ); mov( eax, w );Of course, this code is grossly inefficient and it requires that you declare a couple of temporary variables in your data segment. However, it is very easy to optimize this code by keeping temporary variables, as much as possible, in 80x86 registers. By using 80x86 registers to hold the temporary results this code becomes:
mov( a, eax ); add( b, eax ); mov( y, ebx ); add( z, ebx ); intmul( ebx, eax ); mov( eax, w );x := (y+z) * (ab) / 10;This can be converted to a set of four simple expressions:
Temp1 := (y+z) Temp2 := (ab) Temp1 := Temp1 * Temp2 X := Temp1 / 10You can convert these four simple expressions into the assembly language statements:
mov( y, eax ); // Compute eax = y+z add( z, eax ); mov( a, ebx ); // Compute ebx = ab sub( b, ebx ); imul( ebx, eax ); // This also sign extends eax into edx. idiv( 10, edx:eax ); mov( eax, x );The most important thing to keep in mind is that temporary values, if possible, should be kept in registers. Remember, accessing an 80x86 register is much more efficient than accessing a memory location. Use memory locations to hold temporaries only if you've run out of registers to use.
Ultimately, converting a complex expression to assembly language is little different than solving the expression by hand. Instead of actually computing the result at each stage of the computation, you simply write the assembly code that computes the results. Since you were probably taught to compute only one operation at a time, this means that manual computation works on "simple expressions" that exist in a complex expression. Of course, converting those simple expressions to assembly is fairly trivial. Therefore, anyone who can solve a complex expression by hand can convert it to assembly language following the rules for simple expressions.
10.3.4 Commutative Operators
If "@" represents some operator, that operator is commutative if the following relationship is always true:
(A @ B) = (B @ A)As you saw in the previous section, commutative operators are nice because the order of their operands is immaterial and this lets you rearrange a computation, often making that computation easier or more efficient. Often, rearranging a computation allows you to use fewer temporary variables. Whenever you encounter a commutative operator in an expression, you should always check to see if there is a better sequence you can use to improve the size or speed of your code. The following tables list the commutative and noncommutative operators you typically find in high level languages:
10.4 Logical (Boolean) Expressions
Consider the following expression from a Pascal program:
B := ((X=Y) and (A <= C)) or ((ZA) <> 5);B is a boolean variable and the remaining variables are all integers.
How do we represent boolean variables in assembly language? Although it takes only a single bit to represent a boolean value, most assembly language programmers allocate a whole byte or word for this purpose (as such, HLA also allocates a whole byte for a BOOLEAN variable). With a byte, there are 256 possible values we can use to represent the two values true and false. So which two values (or which two sets of values) do we use to represent these boolean values? Because of the machine's architecture, it's much easier to test for conditions like zero or not zero and positive or negative rather than to test for one of two particular boolean values. Most programmers (and, indeed, some programming languages like "C") choose zero to represent false and anything else to represent true. Some people prefer to represent true and false with one and zero (respectively) and not allow any other values. Others select all one bits ($FFFF_FFFF, $FFFF, or $FF) for true and 0 for false. You could also use a positive value for true and a negative value for false. All these mechanisms have their own advantages and drawbacks.
Using only zero and one to represent false and true offers two very big advantages: (1) The SETcc instructions produce these results so this scheme is compatible with those instructions; (2) the 80x86 logical instructions (AND, OR, XOR and, to a lesser extent, NOT) operate on these values exactly as you would expect. That is, if you have two boolean variables A and B, then the following instructions perform the basic logical operations on these two variables:
// c = a AND b; mov( a, al ); and( b, al ); mov( al, c );// c = a OR b; mov( a, al ); or( b, al ); mov( al, c );// c = a XOR b; mov( a, al ); xor( b, al ); mov( al, c );// b = not a; mov( a, al ); // Note that the NOT instruction does not not( al ); // properly compute al = not al by itself. and( 1, al ); // I.e., (not 0) does not equal one. The AND mov( al, b ); // instruction corrects this problem.mov( a, al ); // Another way to do b = not a; xor( 1, al ); // Inverts bit zero. mov( al, b );Note, as pointed out above, that the NOT instruction will not properly compute logical negation. The bitwise not of zero is $FF and the bitwise not of one is $FE. Neither result is zero or one. However, by ANDing the result with one you get the proper result. Note that you can implement the NOT operation more efficiently using the "xor( 1, ax );" instruction since it only affects the L.O. bit.
As it turns out, using zero for false and anything else for true has a lot of subtle advantages. Specifically, the test for true or false is often implicit in the execution of any logical instruction. However, this mechanism suffers from a very big disadvantage: you cannot use the 80x86 AND, OR, XOR, and NOT instructions to implement the boolean operations of the same name. Consider the two values $55 and $AA. They're both nonzero so they both represent the value true. However, if you logically AND $55 and $AA together using the 80x86 AND instruction, the result is zero. True AND true should produce true, not false. Although you can account for situations like this, it usually requires a few extra instructions and is somewhat less efficient when computing boolean operations.
A system that uses nonzero values to represent true and zero to represent false is an arithmetic logical system. A system that uses the two distinct values like zero and one to represent false and true is called a boolean logical system, or simply a boolean system. You can use either system, as convenient. Consider again the boolean expression:
B := ((X=Y) and (A <= D)) or ((ZA) <> 5);The simple expressions resulting from this expression might be:
mov( x, eax ); cmp( y, eax ); sete( al ); // AL := x = y;mov( a, ebx ); cmp( ebx, d ); setle( bl ); // BL := a <= d; and( al, bl ); // BL := (x=y) and (a <= d);mov( z, eax ); sub( a, eax ); cmp( eax, 5 ); setne( al ); or( bl, al ); // AL := ((X=Y) and (A <= D)) or ((ZA) <> 5); mov( al, b );When working with boolean expressions don't forget the that you might be able to optimize your code by simplifying those boolean expressions. You can use algebraic transformations (especially DeMorgan's theorems) to help reduce the complexity of an expression. In the chapter on lowlevel control structures you'll also see how to use control flow to calculate a boolean result. This is generally quite a bit more efficient than using complete boolean evaluation as the examples in this section teach.
