Re: [PATCH] [4.5 regression] C++ ignores some aligned attributes

[Martin Sebor <msebor@gmail.com>]

On 08/04/2010 08:19 AM, Ulrich Weigand wrote:
> Paul Brook wrote:
>
>> I'm not so sure. Consider the following example:
>>
>> struct JSString
>> {
>>    unsigned char mLength;
>>    static  JSString unitStringTable[];
>> };
>>
>> int var1 =__alignof__ (JSString::unitStringTable);
>>
>> JSString JSString::unitStringTable[] __attribute__ ((aligned (8))) = { 1 };
>> int var2 =__alignof__ (JSString::unitStringTable);
>>
>> int main()
>> {
>>    assert(var1 == var2);
>> }
>>
>> Note that var1 may be in a separate file, and never see the additional
>> attribute.
>>
>> Should we be rejecting this mismatch, rather than papering over what looks
>> like user error?  If the packed attribute can reduce alignment then this is
>> going to cause havoc on STRICT_ALIGN targets.
>
> Well, the only thing my patch does is to revert behavior back to what it was
> in 4.4 and earlier, after having been (clearly inadvertently) changed by an
> unrelated patch.  At least Firefox relies on that behavior; Firefox builds
> would certainly break if we add the error you suggest ...
>
> That said, I have no strong opinion on what the behavior *should* be in
> the case of differing aligned attributes between declaration and definition.
> (However, I would suggest to keep the behavior the same between C and C++
> whereever that makes sense.)

In C++ 0x, the (presumably equivalent) alignof operator had better
yield the same result across all translation units to preserve the
ODR. Consider:

struct S { static int i; };

template <int> struct X { };
template <> struct X<alignof S::i> { enum { e }; };

int S::i [[align(sizeof (int) * 2)]];

int main() {
     return X<alignof S::i>::e;
}

Martin

>
> Bye,
> Ulrich
>