c++debuggingmetaprogramming

Getting rid of #ifndef NDEBUG


Most of my classes have debug variables, and this makes them often look like this:

class A
{
    // stuff
#ifndef NDEBUG
    int check = 0;
#endif
};

and methods might look like this:

for (/* big loop */) {
    // code
#ifndef NDEBUG
    check += x;
#endif
}

assert(check == 100);

Few things are uglier than all those #ifndef NDEBUG's. Unfortunately no compiler I know can optimize the check variable away without these #ifndefs (I don't know if that's even allowed).

So I've tried to come up with a solution that would make my life easier. Here's how it looks now:

#ifndef NDEBUG

#define DEBUG_VAR(T) T

#else

template <typename T>
struct nullclass {
    inline void operator+=(const T&) const {}
    inline const nullclass<T>& operator+(const T&) const { return *this; }
    // more no-op operators...
};

#define DEBUG_VAR(T) nullclass<T>

#endif

So in debug mode, DEBUG_VAR(T) just makes a T. Otherwise it makes a "null class" with only no-ops. And my code would look like this:

class A {
   // stuff
   DEBUG_VAR(int) check;
};

Then I could just use check as if it were a normal variable! Awesome! However, there are still 2 problems that I cannot get solved:

1. It only works with int, float, etc.

The "null class" doesn't have push_back() etc. No biggie. Most debug variables are ints anyway.

2. The "null class" is 1 char wide!!

Every class in C++ is at least 1 char wide. So even in release mode, a class that uses N debug vars will be at least N chars too big. This is in my eyes just unacceptable. It's against the zero-overhead principle which I aim for as much as I can.

So, how do I fix this second problem? Is it even possible to get rid of the #ifndef NDEBUG's without hurting performance in non-debug mode? I accept any good solution, even if it's your darkest C++ wizardry or C++0x.


Solution

  • How about:

    #ifndef NDEBUG
    #define DEBUG_VAR(T) static nullclass<T>
    #endif
    

    Now no additional storage is added to a class where DEBUG_VAR(T) is used as a member, but the declared identifier can still be used as though it were a member.