I don't know why NOT is defined as a function but a macro.
I've heard that function call may be cache-unfriendly.
Will it be better if NOT is a macro？
The only reason I can think of is that a NOT function can be passed as value by adding #'.
not has the semantics of a function: its argument is evaluated perfectly normally (unlike, say, and and so on). It should therefore be a function: macros should be used only for things which cannot semantically be functions.
Making not a function has benefits: simply because it is a function you don't need to write your own version every time you need a function with the semantics of not. Making not a function also has zero cost since the semantics of CL leave a compiler or interpreter entirely free to never actually call the function but rather generate inline code.
Indeed you can see, for instance, in the SBCL sources that this is happening: in src/code/pred.lisp you will find
(defun not (object)
"Return T if X is NIL, otherwise return NIL."
(not object))
You can do this yourself, pretty much:
(declaim (inline neg))
(defun neg (x)
(if x nil t))
(And if you are really feeling your oats define a compiler macro as well, although I can't see any strong reason to do that here.)