MSVC++ variadic macro expansion

Question

So I've got a macro that works nicely in GCC, but not in Microsoft's C++ Compiler. I'm hoping somebody might know of a workaround, or perhaps can explain to me why it behaves this way.

I'm sure this macro isn't exactly "standard", but it would really help me out.

Here is a functional example of the macro:

#define VA_NARGS_IMPL(_1, _2, _3, _4, _5, N, ...) N
#define VA_NARGS(...) VA_NARGS_IMPL(__VA_ARGS__, 5, 4, 3, 2, 1)

#define FULLY_EXPANDED(count, ...) \
  MAC ## count (__VA_ARGS__)

#define SEMI_EXPANDED(count, ...) FULLY_EXPANDED(count, __VA_ARGS__)

#define EXPAND_THESE(...) SEMI_EXPANDED(VA_NARGS(__VA_ARGS__), __VA_ARGS__)

#define ACTUAL_MACRO(x) parent->GetProperty<x>();
#define MAC1(a) ACTUAL_MACRO(a)
#define MAC2(a,b) MAC1(a) ACTUAL_MACRO(b)
#define MAC3(a,b,c) MAC2(a,b) ACTUAL_MACRO(c)
#define MAC4(a,b,c,d) MAC3(a,b,c) ACTUAL_MACRO(d)
#define MAC5(a,b,c,d,e) MAC4(a,b,c,d) ACTUAL_MACRO(e)

Here is how I might use this macro:

struct MyStructure
{
  void Foo()
  {
    EXPAND_THESE(Property1, Property2, Property3, Property4)
  }

  Base * parent;
}

Here's how GCC expands the above:

struct MyStructure
{
  void Foo()
  {
    parent->GetProperty<Property1>(); 
    parent->GetProperty<Property2>(); 
    parent->GetProperty<Property3>(); 
    parent->GetProperty<Property4>();
  }

  Base * parent;
}

But Microsoft for some reason expands all my __VA_ARGS__ as one argument:

struct MyStructure
{
  void Foo()
  {
    parent->GetProperty<Property1, Property2, Property3, Property4>();
  }

  Base * parent;
}

Does anybody know why this is? Is there some trick I can pull to get Microsoft to expand this like GCC? Maybe toss in a couple extra pairs of parentheses?

Macros like this could really help me out in replacing a bunch of "glue" code, but because of this problem, I can't move it into my VS project. Any help would be greatly appreciated!

Thanks.

Answer 1

Coincidentally, I happened to run into this problem just today, and after enough effort I think I've found a solution for my own purposes. The bug is MSVC treats __VA_ARGS__ as a single token in argument lists. But you can work around this by not using it directly within a macro call argument list. This comment suggests the start of an answer to your problems:

#define VA_NARGS(...) VA_NUM_ARGS_IMPL_((__VA_ARGS__, 5,4,3,2,1))
#define VA_NARGS_IMPL_(tuple) VA_NUM_ARGS_IMPL tuple
#define VA_NARGS_IMPL(_1,_2,_3,_4,_5,N,...) N

But then I suspect you'll likely run into the issue of making sure that gets fully expanded to the actual "N" you want, and not to VA_NARGS_IMPL (arg1, arg2, 5, 4, 3, 2, 1), say. I found that my code (which looked like yours) had to change to expand MAC##code all as one unit, and then that had to be separately combined with the argument list. Here's the code that I found worked for me:

#define ASSERT_HELPER1(expr) singleArgumentExpansion(expr)
#define ASSERT_HELPER2(expr, explain) \
   twoArgumentExpansion(expr, explain)

/*
 * Count the number of arguments passed to ASSERT, very carefully
 * tiptoeing around an MSVC bug where it improperly expands __VA_ARGS__ as a
 * single token in argument lists.  See these URLs for details:
 *
 *   http://connect.microsoft.com/VisualStudio/feedback/details/380090/variadic-macro-replacement
 *   http://cplusplus.co.il/2010/07/17/variadic-macro-to-count-number-of-arguments/#comment-644
 */
#define COUNT_ASSERT_ARGS_IMPL2(_1, _2, count, ...) \
   count
#define COUNT_ASSERT_ARGS_IMPL(args) \
   COUNT_ASSERT_ARGS_IMPL2 args
#define COUNT_ASSERT_ARGS(...) \
   COUNT_ASSERT_ARGS_IMPL((__VA_ARGS__, 2, 1, 0))
 /* Pick the right helper macro to invoke. */
#define ASSERT_CHOOSE_HELPER2(count) ASSERT_HELPER##count
#define ASSERT_CHOOSE_HELPER1(count) ASSERT_CHOOSE_HELPER2(count)
#define ASSERT_CHOOSE_HELPER(count) ASSERT_CHOOSE_HELPER1(count)
 /* The actual macro. */
#define ASSERT_GLUE(x, y) x y
#define ASSERT(...) \
   ASSERT_GLUE(ASSERT_CHOOSE_HELPER(COUNT_ASSERT_ARGS(__VA_ARGS__)), \
               (__VA_ARGS__))

int foo()
{
  ASSERT(one); // singleArgumentExpansion(one)
  ASSERT(two, "foopy"); // twoArgumentExpansion(two, "foopy")
}

My mind is too much mush after a few hours solving my own issues to then go and completely solve yours, I'm sorry to say. :-) But I think this is enough to get you to something that works, with a little work.

Answer 2

I know this question is over two years old, but I thought I would try to give a more polished answer to those who still stumble upon this, like I did.

Jeff Walden's answer works and all, but you have to declare FOO_CHOOSE_HELPER/1/2 for each FOO macro you want to have variadic arguments. I have developed a layer of abstraction to solve this issue. Consider the following:

#define GLUE(x, y) x y

#define RETURN_ARG_COUNT(_1_, _2_, _3_, _4_, _5_, count, ...) count
#define EXPAND_ARGS(args) RETURN_ARG_COUNT args
#define COUNT_ARGS_MAX5(...) EXPAND_ARGS((__VA_ARGS__, 5, 4, 3, 2, 1, 0))

#define OVERLOAD_MACRO2(name, count) name##count
#define OVERLOAD_MACRO1(name, count) OVERLOAD_MACRO2(name, count)
#define OVERLOAD_MACRO(name, count) OVERLOAD_MACRO1(name, count)

#define CALL_OVERLOAD(name, ...) GLUE(OVERLOAD_MACRO(name, COUNT_ARGS_MAX5(__VA_ARGS__)), (__VA_ARGS__))

With this architecture you can define variadic macros as such:

#define ERROR1(title) printf("Error: %s\n", title)
#define ERROR2(title, message)\
    ERROR1(title);\
    printf("Message: %s\n", message)
#define ERROR(...) CALL_OVERLOAD(ERROR, __VA_ARGS__)

#define ASSERT1(expr) singleArgumentExpansion(expr)
#define ASSERT2(expr, explain) twoArgumentExpansion(expr, explain)
#define ASSERT(...) CALL_OVERLOAD(ASSERT, __VA_ARGS__)

With Jeff's answer you would have to define the macros as follows:

#define ERROR1(title) printf("Error: %s\n", title)
#define ERROR2(title, message)\
    ERROR1(title);\
    printf("Message: %s\n", message)

#define ERROR_CHOOSE_HELPER2(count) ERROR##count
#define ERROR_CHOOSE_HELPER1(count) ERROR_CHOOSE_HELPER2(count)
#define ERROR_CHOOSE_HELPER(count) ERROR_CHOOSE_HELPER1(count)

#define ERROR(...) GLUE(ERROR_CHOOSE_HELPER(COUNT_ARGS_MAX5(__VA_ARGS__)),\
    (__VA_ARGS__))

#define ASSERT1(expr) singleArgumentExpansion(expr)
#define ASSERT2(expr, explain) twoArgumentExpansion(expr, explain)

#define ASSERT_CHOOSE_HELPER2(count) ASSERT##count
#define ASSERT_CHOOSE_HELPER1(count) ASSERT_CHOOSE_HELPER2(count)
#define ASSERT_CHOOSE_HELPER(count) ASSERT_CHOOSE_HELPER1(count)

#define ASSERT(...) GLUE(ASSERT_CHOOSE_HELPER(COUNT_ARGS_MAX5(__VA_ARGS__)),\
    (__VA_ARGS__))

It's not a big deal, however I like my code to be as concise as possible. It also helps exponentially, if you are using several variadic macros, to reduce code duplication and the complications that can cause. As far as I know, this method is also portable. I have tested it on many of the most common compilers and they produced the same results.

Example use:

int foo()
{
    ASSERT(one); // singleArgumentExpansion(one)
    ASSERT(two, "foopy"); // twoArgumentExpansion(two, "foopy")

    ERROR("Only print a title");
    ERROR("Error title", "Extended error description");
}

Answer 3

Microsoft has rewritten the C/C++ preprocessor, but it is not enabled by default for "backwards-compatibility", i.e. they prefer bug-compatibility with their own products instead of portability or standard compliance.

It seems that you can fix the __VA_ARGS__ handling by adding a /experimental:preprocessor flag to the command line.