On Wed, Dec 11, 2013 at 12:28:32PM -0700, Eric Blake wrote:

On 12/11/2013 12:15 PM, Eric Blake wrote:

struct _virObjectEvent {
    virObject parent;
    int eventID;
    virObjectMeta meta;
};

Only has alignment specified by virObject (which in turn is unsigned
int, int, void*),

struct _virObject {
    unsigned int magic;
    int refs;
    virClassPtr klass;
};

I think one possible solution would be as simple as altering
src/util/virobject.h to change 'magic' from 'unsigned int' to 'unsigned
long long' - then ALL virObject structs will be forcefully aligned to
the worst case between void* and long long, so that any subclass can use
long long without requiring stricter alignment than the parent class,
and so that downcasting code like domain_event.c no longer warns.  But
it does make every object consume more memory on 64-bit platforms (from
16 bytes into 24 bytes), is that okay?

Or maybe even change _virObject to contain a union:

struct _virObject {
    union {
        long long align;
        struct {
            unsigned int magic;
            int refs;
        } s;
    } u;
    virClassPtr klass;
}

which keeps the size at 16 bytes on 64-bit platform, keeps things at 12
bytes on 32-bit platforms that don't care about long long alignment, and
for ARM (*) would change things from 12 to 16 bytes with 8-byte
alignment for the long long.

Yeah, that means using obj->u.s.refs instead of obj->refs, but most code
shouldn't be directly mucking with object-internal fields, so hopefully
the fallout would be limited to just virobject.c (if only C99 had
anonymous struct/unions; C11 does, but we don't require that yet).

(*) Am I correct that your platform with the compile failure is a 32-bit
ARM platform with 4-byte pointers? Because if it were 64-bit, then I
would have guessed that an 8-byte pointer would already be forcing
8-byte alignment, such that use of 'long long' in a subclass wouldn't be
warning about changed alignment.

That seems reasonable to me - it makes sense that we should have
our base object type be nicely aligned, instead of trying to fix
this in the events code (and potentially anywhere else using
objects in the future).

Daniel