On Wed, May 22, 2013 at 3:57 PM, Sandra Loosemore
<san...@codesourcery.com> wrote:
> On 05/22/2013 02:01 AM, Richard Biener wrote:
>>
>> On Wed, May 22, 2013 at 3:57 AM, David Edelsohn <dje....@gmail.com> wrote:
>>>
>>>
>>> Increasing the alignment of arrays within structs and unions would be
>>> nice, but that probably will change the ABI. I think that they best we
>>> may be able to do is increase the alignment if the array is the first
>>> element of the struct or union, see ROUND_TYPE_ALIGN for AIX.
>>> Although this might be more trouble than it is worth.
>>
>>
>> Maybe the idea was to increase the alignment of the struct (without
>> affecting it's layout) when that increases the alignment of a contained
>> array member. Like for
>>
>> struct { int i; int j; float a[1024]; int x; };
>>
>> where aligning to sizeof (int) * 2 would get 'a' a bigger alignment.
>
>
> In fact, this is what the patch I posted earlier this week does, except the
> alignment it looks for is 16 bytes. In other words, it doesn't change the
> alignment of arrays inside structs (which would cause all sorts of
> compatibility problems), but it checks to see whether there is already some
> array inside the struct aligned on a 16-byte offset so that the whole
> containing struct object would benefit from being aligned on a 16-byte
> boundary.
>
> This was a generalization of the suggestion from the previous review to
> align structs that contain an array as the first or only element.... I
> realized we'd want to handle unions too, and look at all fields in a union
> instead of just the first one in a struct, and unifying the two cases made
> the code come out tidier. :-)
Right. Btw, the code should all be target independent ... just the desired
alignment for an array should be queried from it. pass_ipa_increase_alignment
and/or the vectorizer code itself should use that info.
Richard.
> -Sandra
>
