Jens Gustedt's Blog

February 2, 2011

Handling control flow inside macros

Filed under: C99, language, P99, preprocessor, syntax — Jens Gustedt @ 20:35

When people program macros that they want to be executed anywhere where a statement could be, they often use the

do {...} while(0)

construct. This construct has a big disadvantage in that it may change control flow in an unexpected way when you’d use it as a generic macro tool to collect statements:

#define PXX_BLOCK(...) do { __VA_ARG__ } while(0)
.
#define PXX_HYPOTHETIC_BREAK(X)           \
PXX_BLOCK(                                \
if ((X)%2) {                              \
     printf("some thing is wrong:\n");    \
     break;                               \
})
.
for (size_t i = 0; i < n; i+=2) {
    call_a_function(&i);
    PXX_HYPOTHETIC_BREAK(i);
}

In this example the break statement of the macro will not do what the user of PXX_HYPOTHETIC_BREAK might expect, namely to break out of the for-loop, but just break out of the do-while and continue executing the for:

for (size_t i = 0; i < n; i+=2) {
    call_a_function(&i);
    do {
      if ((X)%2) {
         printf("some thing is wrong:\n"); 
         break;
      } while (0);
   }
}

P99 uses another approach to do such statement grouping inside macros that does not present surprises concerning the control flow:

#define P99_NOP ((void)0)
#define P99_PREFER(...) if (1) { __VA_ARGS__ } else
#define P99_BLOCK(...) P99_PREFER(__VA_ARGS__) P99_NOP

and with this replacement of the block macro we obtain as expansion for PXX_HYPOTHETIC_BREAK(i):

if (1) {
  if (i%2) {
    printf("some thing is wrong:\n");
    break;
  }
} else ((void)0)
  • This is placed perfectly fine where C expects a single statement and not a block.
  • The break statement in there is not hidden by a loop statement and has its effect on the surrounding for loop.
  • The if statement has a corresponding else, so an eventual other else of the user code will not bind to this one here.
  • The else and its depending statement ((void)0) is never executed. But even if it where a constant that is cast to void is an expression without side effects that is eliminated by the compiler.

Other macros that P99 derives from this:

#define P99_AVOID P99_PREFER(/* NOP */)
#define P99_XCASE P99_AVOID case
#define P99_XDEFAULT P99_AVOID default

The first may at first serve as some sort of comment marker.

P99_AVOID dont_do_that_thing(x, y, z);

Will never call the function dont_do_that_thing since this statement is unreachable. But other than a comment marker // this will always check if the call is syntactically correct. If other things change in your code that would make this statement invalid, you’d still notice and be forced to keep your code up-to-date.

For the two other macros look at a hypothetical usage of these “X” cases, X for eXclusive:

switch (c) {
  P99_XCASE 5: print("take 5\n");
  P99_XCASE 7: return;
  P99_XDEFAULT: ++i;
}

The idea is here that these are all exclusive cases, no need to even put break after the cases to stop the switch.

switch (c) {
  if (1) { } else case 5: print("take 5\n");
  if (1) { } else case 7: return;
  if (1) { } else default: ++i;
}

Here all the if (1) have the effect than the statement that is depending on the else clause will never be executed unless we jump directly into it via the case label. (For those that have doubts, the resulting code is valid C. The switch statement is much more generic than you think.)

Another application is to hide some variables. Your remember perhaps my entry about scope bound resource management with for scopes. The idea there is to use the variable of a for-scope as a local variable for a block or statement but that is still only executed once. One drawback of using a plain for-variable is that C99 doesn’t allow these to be static.

for (int p00 = 1; p00; p00 = 0)
  for (pthread_mutex_t* point = 0; p00; p00 = 0)
    P99_PREFER(
              static pthread_mutex_t crit = PTHREAD_MUTEX_INITIALIZER;
              crit.lock();
              point = &crit;
              goto JUMPLABEL;
             ) 
       JUMPLABEL: for (; p00; (point->unlock(), p00 = 0))

This has a mutex “crit” that is locked on entry and unlocked when leaving. P99 gives some more programming tools at hand to write such things in macros using P99_GUARDED_BLOCK for the lock/unlock calls and P00_BLK_DECL_STATIC to encapsulate this kind of static declaration as we have seen it here. A simple to use macro CRITICAL_SECTION can be defined like this:

#define MUTUAL_EXCLUDE(MUT)                                    \
P99_GUARDED_BLOCK(pthread_mutex_t*,                            \
      P99_FILEID(mut),                                         \
      &(MUT),                                                  \
      pthread_mutex_lock(P99_FILEID(mut)),                     \
      pthread_mutex_unlock(P99_FILEID(mut)))

#define CRITICAL_SECTION                                                    \
P00_BLK_START                                                               \
P00_BLK_DECL_STATIC(pthread_mutex_t, orwl__crit, PTHREAD_MUTEX_INITIALIZER) \
MUTUAL_EXCLUDE((*orwl__crit))
.
.
CRITICAL_SECTION {
  /* do something critical and important here */
  a += 42;
}
Advertisements

2 Comments

  1. I’ve noticed that you generally use the style of putting the case blocks in {…} brackets. That’s not my preferred style, and

    switch (c) {
        P99_XCASE(5):
            do_something();
        P99_XCASE(7):
            do_something_else();
            something_yet_else();
    }
    

    will break; it will resolve to this (indenting added for clarity):

    switch (c) {
        if (1) { }
        else
            case 5: do_something();
        if (1) { }
        else
            case 7: do_something_else();
        something_yet_else();
    }
    

    Notice that for c == 5, both lines do_something(); and something_yet_else(); will be executed, which is probably not what the user wants.

    Would you please explicitly document this limitation?

    Comment by Joel Salomon — February 9, 2011 @ 21:13

    • Right, good to know that this causes confusion. For me it looked natural: the XCASE prefixes the statement or block that is exclusive. I also have no spontaneous idea, how I would implement such an exclusive case that extends up to the beginning of the next case or the end of the switch.

      Yes, I’ll improve the documentation with this in view.

      Comment by Jens Gustedt — February 9, 2011 @ 21:29


RSS feed for comments on this post.

Sorry, the comment form is closed at this time.

Blog at WordPress.com.