Currently,
BOOST_VARIADIC_SIZE()
evaluates to 1.
In the preprocessor, variadic arguments are always one or more. If you need zero or more arguments instead should use a tuple or a sequence. I'm not sure that Boost supports zero-sized tuples or sequences, yet.
Based on http://gustedt.wordpress.com/2010/06/08/detect-empty-macro-arguments/ it can be improved to return 0 for an empty argument list.
Heres an excerpt about tuples from Chaos PP documentation(I'm not sure there's a link on the web for it), that explains it well(perhaps we should adapt this to boost docs since this question comes up often): Several problems arise from this situation when using tuples as data structures. The first is that there is no general way of detecting the difference between () and (1). It is possible to get fairly close in various ways by restricting the types of elements, but a general solution is impossible. Even if such detection **was** possible, another problem surfaces. Consider again the (,,) tuple, and consider an algorithm that operates on each element of a tuple. If the algorithm was to attempt to individually process elements one- by-one, it might use CHAOS_PP_TUPLE_HEAD and CHAOS_PP_TUPLE_TAIL. Everything is fine for CHAOS_PP_TUPLE_TAIL((,,)) (which yields (,)). However, CHAOS_PP_TUPLE_TAIL((,)) yields () (i.e. the theoretical nil tuple), and an element has been silently skipped. The underlying problem is that emptiness (or "nothing") is a useful and perfectly valid form of data in its own right. As a consequence, algorithms and primitives that operate on tuples algorithmically need to use a distinct representation for nil that has no ambiguity. This representation is CHAOS_PP_TUPLE_NIL() which simply expands to nothing. This model always treats () as a unary tuple, and CHAOS_PP_TUPLE_TAIL(()) expands to nothing. Paul Fultz II