Fix alignment computation at exapnsion time

[Richard Sandiford]

Andrew Pinski <pinskia@physics.uc.edu> writes:
> On Jul 1, 2006, at 4:24 PM, Jan Hubicka wrote:
>
>>
>> Hi,
>> in earlier unit-at-a-time the variables was assembled first,  
>> function bodies later.
>> This order was changed to allow more optimizations, but now we  
>> forget to compute the
>> alignment of variable early having overly conservative assumptions  
>> at a time we expand
>> functions.
>
> Hmm, does this fix PR 27770 also?

No, at least not in general.  See the description of the problem in the PR.
The vectoriser is increasing the alignment of a variable beyond that normally
imposed by assemble_variable, and the vectoriser is doing this after we've
used the old alignment in a previous function.

Richard