[C++ PATCH] [PR19163] Fix attribute aligned on dependent arraytypes

[Mark Mitchell]

Giovanni Bajo wrote:
> Mark Mitchell <mark@codesourcery.com> wrote:
> 
> 
>>Second, I'm not sure that you've dealt with RTH's question fully.  In
>>the case that the original type is dependent, does TYPE_ALIGN (t)
>>have a sensible value?
> 
> 
> If TYPE_USER_ALIGN is true, the answer is yes, and TYPE_ALIGN carries the
> alignment specified by the user in the attribute. If TYPE_USER_ALIGN is no, I
> do not know for sure: probably it does not carry any useful information since
> the type was not layed out. But my (latest) patch only commits an action if
> TYPE_USER_ALING is true.

I don't think that we'll be handling the dependent case correctly.

Consider:

template <typename T>
struct S {
   T t;
} __attribute__((aligned (1)));

template struct S<double>;

I looked at this in the debugger, and TYPE_USER_ALIGN looks like its 
being set for S<T>.  So, I would expect that with your patch TYPE_ALIGN 
would be set to 1 byte for S<double>.

As you say:

>>And the documentation says that the "aligned" attribute
>>can only increase alignment; how do we know that this constraint is
>>being applied?
> 
> 
> I did not know about this constraint. 

It turns out that (apparently) you don't get an error; it's just that 
your alignment request is ignored.

> I do not know where it is enforced for
> non-tempalte type/decls in the first place, I do not see such a check in
> handle_align_attribute. Probably it's deeper in the middle-end, so we should
> not care in the FE.

It might be that what you say is true.  If so, you need to find that 
place, and/or verify that the alignment of the example above is the same 
as without the alignment attribute; it should be ignored.  If that works 
OK, please let me know; I'll reconsider the patch.

Thanks,

-- 
Mark Mitchell
CodeSourcery, LLC
mark@codesourcery.com
(916) 791-8304