Some of the least readable code I have seen has been code where every
trick in the book was used to make it as short as possible, so using
extreme optimisaton tricks is probably a much better obfuscator than
inserting random code -- even random control-structure code.

Some common optimisations do, indeed, make the code less readable:
Strength reduction, loop scheduling, array blocking, common
subexpression elimination and so on.  So using these aggressively
would work well.  But there will be code where such optimsations do
not apply.  Here, you could use some of the following tricks:

 - Split a variable x into two variables p and q, such that x=p+q.
   Modifications to x are made into modifications of p _or_ q.  Uses of
   x (except those for self-modification such as x=x+1) are replaced by
   uses of p+q.  This will make the code somewhat larger and slower, but
   not by much.  Common subexpression elimination can remove some of the
   p+q additions without increasing readability.

 - Replace two otherwise unrelated variables x and y by p and q such
   that p=x+y and q=x-y (so x==(p+q)/2 and y==(p-q)/2).

 - A variable x holding small non-negative integers is replaced by a
   variable p = 2^x.  Additions and subtractions of constants to x are
   made into shifts.  Using x is more tricky, though, unlesss there is a
   "count leading zeroes" instruction.

        Torben