On 4/27/22 13:02, Joseph Myers wrote:
> On Wed, 27 Apr 2022, Jason Merrill via Gcc-patches wrote:
> 
>> +      if (typedef_variant_p (type))
>> +	{
>> +	  /* Set up the typedef all over again.  */
> 
> This seems wrong when the typedef is just being used in another
> declaration with the mode attribute, as opposed to being defined using the
> mode attribute.  E.g. the following test is valid and accepted before the
> patch, but wrongly rejected after the patch because the typedef has had
> its type changed.
> 
> typedef int I;
> int x;
> I y __attribute__ ((mode(QI)));
> extern I x;

Ah, good point.  Fixed thus: