Allow generate_element_init in tree-sra.c to fail

[Andrew Pinski]

> 
> > 
> >     Well it looks like we can no longer call scalarize_init with NULL, as
> >     the comment says before that function, we treat NULL as an empty
> >     constructor.
> > 
> > It's precisely because we treat it as empty that this test has to be there
> > because a null DECL_INITIAL doesn't mean that.
> 
> Doesn't empty really mean filled with zero?

If so then the check for NULL you added instead of error_mark_node is wrong
as we no longer replace empty (NULL) constructors which means you just introduced
another regression.

> 
> Andrew
> 
> 
>