Re: [boost] Re: New version of valunion and stack_allocator uploaded

From: Douglas Gregor (gregod_at_[hidden])
Date: 2002-03-01 10:09:01

On Friday 01 March 2002 09:06 am, you wrote:
> > The type to align against is a union, where the type 'X' of each member
> > is
>
> a
>
> > type 'U' if sizeof(U) <= sizeof(type_we_want_to_make_storage_for) and
> > char otherwise.
>
> So if the type I'm storing is char[6], it will over-align, yes?

Yes. That's not a fault of the approach, but of the implementation.

> In my application it's much more important to actually produce an
> appropriate alignment than that the size is minimal, but that won't be true
> for many. Since this is all a crapshoot anyway and the language gives no
> guarantees of success it seems like Brian's algorithm ought at least to
> check that sizeof(U) <= alignment_of(type_we_want_to_make_storage_for).
>
> -Dave

Ok, but I still like Brian Parker's design better, and it's easy to change 
the algorithm. Here's his code, that I slightly modified to do as you 
mentioned above:

        template<typename T, size_t sz, size_t align>
        struct member {
          typedef typename ct_if<sizeof(T) <= sz && sizeof(U) <= align,
                                  T,
                                  char>::type type;
        };

        template<size_t sz, size_t align>
        union force_align {
                typename member<short, sz, align>::type a;
                typename member<int, sz, align>::type b;
                typename member<long, sz, align>::type c;
                typename member<float, sz, align>::type d;
                typename member<double, sz, align>::type e;
                typename member<long double, sz, align>::type f;
                typename member<void*, sz, align>::type g;
                typename member<ALIGN_FPTR, sz, align>::type h;
                typename member<POD, sz, align>::type i;

                // todo: add any other built-in or POD types with stricter
                // alignment restrictions here
        };

No typelist, no compile-time for_each, and only one trivial metaprogramming 
construct used (ct_if).

        Doug