Typically, at the assembly level, doing a "compare" will change the contents of some CPU flags, but not registers. Normally, those CPU flags would then influenc

Typically, at the assembly level, doing a "compare" will change the contents of some CPU flags, but not registers. Normally, those CPU flags would then influence future instructions - most often a "GOTO if this flag is set" instruction. In this case, you want to return the result of the comparison from a function. Since functions return values in registers, you'd need to do some extra work to copy the result into a register from the flag. Between compiler optimisations, and the cost of your function call in the grand scheme of things, the difference is likely negligible. This may, or may not, have been the case in 1991. Also, casting to an integer pointer rather than a char pointer means the code has to divide by 4 at least once. (He did write this "optimisation" as an undergraduate!) EDIT: Actual assembly comparison, from gcc -O3 (Apple's 4.2.x). subl (subtract) and cmpl (compare) take the same amount of time. But then _compare has to do setne to to store the "not equal" flag into a register, then movzbl to expand that result from 8 bits (which is what setne writes) to the full size of an int. _subtract: // return (int)((char*)key1 - (char*)key2)) pushl %ebp movl %esp, %ebp movl 8(%ebp), %eax subl 12(%ebp), %eax leave ret _compare: // return key1 != key2 pushl %ebp movl %esp, %ebp movl 12(%ebp), %eax cmpl %eax, 8(%ebp) setne %al movzbl %al, %eax leave ret