On Jul 6, 2012, at 2:31 PM, Eli Friedman wrote:

> On Fri, Jul 6, 2012 at 2:31 PM, Chad Rosier <[email protected]> wrote:
>> 
>> On Jul 6, 2012, at 1:45 PM, Eli Friedman wrote:
>> 
>>> On Fri, Jul 6, 2012 at 1:14 PM, Chad Rosier <[email protected]> wrote:
>>>> All,
>>>> Attached is a patch that enables a per target max vector alignment field 
>>>> (e.g., 32-byte alignment for x86 due to AVX).  Currently, if no aligned 
>>>> attribute is specified the alignment of a vector is inferred from its 
>>>> size.  Thus, very large vectors will be over-aligned with no benefit.  
>>>> With this patch in place the alignment will be the lesser of the size of 
>>>> the vector and the target max alignment unless an aligned attribute is 
>>>> used.
>>> 
>>> It doesn't seem likely that x86 itself will go beyond 32-byte vectors,
>>> but this sort of thing is hard to predict; a few years ago, it wasn't
>>> obvious that x86 was going to introduce vectors with 32-byte
>>> alignment.
>> 
>> Ok.  Do you have opinion for or against the patch or were you just making a 
>> general comment?
> 
> I'm not entirely sure whether that's an argument for or against this
> patch, just that we should be careful here; it's not uncommon for
> architectures to introduce new vector types, and they usually want to
> be natively aligned.

Understandable;  my assumption was that if such a case occurs (i.e., a new 
vector type is introduced with a size larger then the current target max), then 
we would increase the max target alignment once the vector extensions are added.

> -Eli

_______________________________________________
cfe-commits mailing list
[email protected]
http://lists.cs.uiuc.edu/mailman/listinfo/cfe-commits

Reply via email to