C/C++ most efficient if statement evaluation

Question

X is true nearly 99.9% of the time but I need to handle Y and Z as well. Although the body of the X condition is empty, I'm thinking it should be faster than potentially checking 2 other conditions Y and Z if the X condition is omitted. What do you think?

if (likely(X))
{
}
else if (unlikely(Y))
{
...
}
else if (unlikely(Z))
{
...
}

The answer depends on your logic. Does condition X *always* imply not-Y and not-Z? — Greg Hewgill, Jan 19 '12 at 21:48
@PaulR was the question "do i need to worry about this?" or "what's the fastest way to do this?" — d7samurai, Mar 06 '15 at 01:46

log0 · Accepted Answer · 2012-01-20T07:28:27.967

11

You might want to know what exactly happens when you use likely or unlikely:
http://kerneltrap.org/node/4705

I would personally write

if (unlikely(!X))
{
  if (unlikely(Y))
  {
  ...
  }
  else if (unlikely(Z))
  {
   ...
  }
}

Which means if x, continue execution, else jump to if body.

edited Jan 20 '12 at 07:28

answered Jan 19 '12 at 22:03

log0

10,489
4
28
62

score 8 · Answer 2 · answered Jan 19 '12 at 21:49

8

As usual, when in doubt profile; anyhow, if I were to read that code I would find much clearer something like:

if (!likely(X))
{
    if (unlikely(Y))
    {
    ...
    }
    else if (unlikely(Z))
    {
    ...
    }
}

answered Jan 19 '12 at 21:49

Matteo Italia

123,740
17
206
299

1

I am not really sure that the negation of unlikely will work as expected ?? I would rather write `unlikely(!x)`. This means that there will be no jump if `x` is false. – log0 Jan 19 '12 at 21:58
@Ugo I like the solution posted by Matteo and I rewrote it using your unlikely(!x) instead. – chriskirk Jan 19 '12 at 22:01
@Ugo: I assumed that `likely(X)` was the whole condition to be evaluated, as in his pseudocode... still, the point I'm making should be clear. – Matteo Italia Jan 19 '12 at 22:15
2

@MatteoItalia: `likely` and `unlikely` are macros provided by GCC that wraps compiler specific builtins that *hint* to the compiler whether a condition is likely or not. (gcc's is `__builtin_expect` I think) – Matthieu M. Jan 20 '12 at 07:25

score 1 · Answer 3 · answered Jan 19 '12 at 22:09

If a compiler is to assume, it would generally favor the first condition true.

you can use something like __builtin_expect to control this (as Ugo detailed).

If it's in a loop, you should measure because hardware is also a consideration -- it's not just the source and the compiler. How's your cache, and how's branch prediction working for you? Profile. Alter. Profile. Compare.

MikeFHay · Answer 4 · 2012-09-09T12:42:57.743

0

Do you really need the unlikely's on y and z? It sounds from your question like if x is not true, then y or z must be true, in which case those unlikely's are incorrect and I would go with this:

if (unlikely(!X))
{
  if (Y)
  {
  ...
  }
  else //Z must be true
  {
   ...
  }
}

(Would have made this a comment on Ugo's answer, but I don't have enough reputation to leave comments.)

edited Sep 09 '12 at 12:42

answered Aug 10 '12 at 10:34

MikeFHay

8,562
4
31
52

C/C++ most efficient if statement evaluation

4 Answers4