Portable branch prediction hints

The canonical way to do static branch prediction is that if is predicted not-branched (i.e. every if clause is executed, not else), and loops and backward-gotos are taken. So, don't put the common case in else if you expect static prediction to be significant. Getting around an untaken loop isn't as easy; I've never tried but I suppose putting it an an else clause should work pretty portably.

Many compilers support some form of #pragma unroll, but it will still be necessary to guard it with some kind of #if to protect other compilers.

Branch prediction hints can theoretically express a complete description of how to transform a program's flow-control graph and arrange the basic blocks in executable memory… so there are a variety of things to express, and most won't be very portable.

As GNU recommends in the documentation for __builtin_expect, profile-guided optimization is superior to hints, and with less effort.


In most cases, the following code

if (a)
{
   ...
}
else
{
    ...
}

is actually

evaluate(A)

if (!A)
{
   jmp p1
}

... code A

   jmp p2

p1:

... code !A

p2:

Note that if A is true, "code A" is already in the pipeline. The processor will see the "jmp p2" command ahead, and will load p2 code to the pipeline.

If A is false, the "code !A" may not be in the pipleline, therefore it may be slower.

Conclusions:

  1. do If(X) if X is more likely than !X
  2. try to evaluate A as early as possible, so that the CPU can dynmically optimize the pipeline.

:

evaluate(A)

do more stuff

if (A)
   ...

Optimization is inherently a compiler thing, so you have to use compiler functionality to help it. The language itself doesn't care about (or mandate) optimizations.

So the best you can do without compiler-specific extensions is organize your code in such a way where your compilers will "do the right thing" without help. But if you want to be sure, tap in to compiler extensions. (You might try abstracting them behind the preprocessor, so your code remains portable.)