[boost] RE: type-traits update [is_convertible & alignment_of]

From: John Maddock (John_Maddock_at_[hidden])
Date: 2000-08-01 05:59:31

Steve,

>This also screens out any empty UDTs that are convertible to int -- a
>ridiculous situation, but it should be doc'ed, IMO.

Yes I know, that's why I haven't used it for is_enum.  Actually I can't
really imagine an empty class that could usefully be converted to int, but
as you say I'll document it.

>Sorry!  I'll try to be gentle this time: it's still not guaranteed to
work.
>;)  However, I think it is good enough -- it would work for all "normal"
>compilers (all I know of).

>The compiler can also add padding to the end of its struct, although in
your
>implementation I cannot imagine a compiler wanting to do so (since T is
>already padded).

Yes, the compiler can add as much extraneous padding as it wants, however
as I said the result must always be a multiple of the actual alignment (but
not necessarily a multiple of 1).  The reasoning follows what happens if
you have an array of alignment_of<T>::padded, the address of each
sub-object of type T in the array is given by:

base + offsetof(padded::t) + n*sizeof(padded)

In order for each sub-object in the array to be aligned it is necessary
that both offsetof(padded::t) be a multiple of T's alignment, and for
n*sizeof(padded) to be a multiple of the correct alignment for all n.  

From this it follows that sizeof(padded) is a multiple of the alignment of
T. 

Further, sizeof(padded)>sizeof(T), for all T, so:

sizeof(padded) - sizeof(T) == k1*Align(T) - k2*Align(T) : for some integers
k1, and k2 with k1 > k2.

hence sizeof(padded) - sizeof(T) == (k1-k2) * Align(T) : with k1-k2 >= 1

I hope this makes some kind of sence, formal maths isn't my strong point
(in case you hadn't noticed!).

- John.