Why bottom test loop is preferable?

Question

I heard someone once say that compilers frequently move the loop conditions to the bottom of the loop. That is, loops like these:

while (condition) {
    ...
}

is changed to :

if (condition) {
     do {
         ...
     } while (condition);
}

regarding machine independent optimization, why is latter preferable?

Answer 1

Without compiler optimisation, the first loop goes to assembly code like this:

  @@:
cmp ... ; or test ...
jz @f

...
jmp @b

Whereas the second loop goes to something like this:

jmp bottom

  @@:
...

  bottom:
cmp ... ; or test ...
jz @b

Conditional jumps are usually predicted to be taken so the first method could potentially lead to more pipeline/instruction cache flushes.

However, the most important thing is that for the first loop, there are two branches available per loop iteration (2N), whereas in the second, each loop iteration only has one branch with a fixed overhead of the first unconditional jump (N+1).

For more information on loop optimisation, see page 88 of this assembly optimisation guide.