bubble sort program

Each time you call order, you need to call orderab 1 fewer time.  In addition, you could put orderab inline.  I left main unrolloed since it is still called a constant number of times.  Also, since you have dwords, shouldn't your indexes go by 4?

Code segment
     Jmp  main
     Add1 dw 9,3,2,7,1,4,2,5,8,1

Main:     mov cx, 9
     call order
     mov cx, 8
     call order
     mov cx, 7
     call order
     mov cx, 6
     call order
     mov cx, 5
     call order
     mov cx, 4
     call order
     mov cx, 3
     call order
     mov cx, 2
     call order
     mov cx, 1
     call order
     ret

order:     xor si, si
orderab:     mov ax,add1[si]
     cmp ax,add1[si+4]
     jle finish
     mov bx,add1[si+4]
     mov add1[si],bx
     mov add1[si+4],ax
finish:     add si, 4
     loop orderab
     ret

This will reduce the work almost in half, but add some overhad from the loop.  If you are concerned about average run-time, you could do the following.  It will run very quickly if the list is already in order (just 9 compares!), but slower than the above code if it is in reverse order (same number of compares, with more overhead).

Main:  mov cx, 9
order:     xor si, si
     mov di, cx
     xor dx, dx
orderab:
     mov ax,add1[si]
     cmp ax,add1[si+4]
     jle finish
     mov dx, di
     sub dx, ci
     mov bx,add1[si+4]
     mov add1[si],bx
     mov add1[si+4],ax
finish:     add si, 4
     loop orderab
     inc dx
     mov cx, dx
     loop order
     ret

The best way to make the sort more efficient is to use a different algorithm such as quicksort, heapsort, or mergesort.  Those types of sorts require  n* log n comparisons on average, while a sort such as bubble sort requires n*n comparisons.  These algorithms are a lot faster, especially for arrays larger than 300 or so.  For some examples of other algorithms and source code, try http://www.cs.ubc.ca/spider/harrison/Java/sorting-demo.html.

Actually, an optimized bubble-sort will require only n comparisons on an array which is already sorted.  Quicksort, heapsort, mergesort and insertionsort all take n*log n.  If your data set is already mostly sorted, bubble-sort will give you the best average runtime.

if the list is sorted and all you are trying to do is insert an element you dont need anywhere near n comparisons, you will only need log n

each comparison you are chopping the list in 1/2

Sorting a list and inserting an element into a sorted list are two different problems.

There are a few algorithms for inserting a new element into a sorted list.  There is a binary search insertion which requires log (n) compares, and there is a linear search insertion which can take from 1 to n compares.  If your data is in a linked list or you can start off taking a guess as to about where a piece of data should go, the linear search can be faster on average.

Dispite the big-O analysis, sometimes a generally "slower" algorithm can be faster when you take into account your specific data-set.