...

Do I understand correctly that you want the memset call to be skipped for those compiler versions that have implemented value-initialization entirely correctly?

Peter Foelsche wrote:

I'm only concerned with boost::mpl::for_each This function calls a function object by passing an object of the matching type. The contents of this object usually do not matter -- and in case they do, the programmer can write a default constructor to initialize this object. So there is no need in boost::mpl::for_each to call value_initialized.

So you think the issue can be resolved properly, merely by modifying mpl/for_each.hpp, right? You proposed at http://lists.boost.org/Archives/boost/2009/09/155805.php to modify mpl/for_each.hpp by changing: ! value_initialized<arg> x; ! aux::unwrap(f, 0)(boost::get(x)); Into: ! aux::unwrap(f, 0)(arg()); I see some subtle differences here. First of all, boost::get(x) return a reference to an lvalue, while arg() is an rvalue. Do you think that's okay? Secondly, things *might* go wrong if arg doesn't have a user-declared default-constructor *and* the compiler hasn't correctly implemented value-initialization. That's why Fernando and I added the memset call to boost::value_initialized, of course. But you think the data members of the temporary arg() won't be accessed anyway, right? Unless arg has a user-declared default-constructor? I'm sorry I don't know mpl/for_each.hpp well enough. But if the answer to each of the above questions is yes, I guess your proposal is fine. Kind regards, Niels -- Niels Dekker http://www.xs4all.nl/~nd/dekkerware Scientific programmer at LKEB, Leiden University Medical Center