Re: simplify regular expression locale global variables

On 15.10.24 12:08, Alvaro Herrera wrote:
> On 2024-Oct-15, Peter Eisentraut wrote:
> 
>> @@ -253,8 +241,9 @@ pg_set_regex_collation(Oid collation)
>>   		 * catalog access is available, so we can't call
>>   		 * pg_newlocale_from_collation().
>>   		 */
>> +		static struct pg_locale_struct dummy_locale = {.ctype_is_c = true};
>> +
>> +		locale = &dummy_locale;
>>   	}
>>   	else
>>   	{
>> @@ -264,121 +253,80 @@ pg_set_regex_collation(Oid collation)
>>   			ereport(ERROR,
>>   					(errcode(ERRCODE_FEATURE_NOT_SUPPORTED),
>>   					 errmsg("nondeterministic collations are not supported for regular expressions")));
>> [...]
>>   	}
>>   
>>   	pg_regex_locale = locale;
>>   }
> 
> Hmm, is it valid to make pg_regex_locale point to a function-local
> static here?  The lifetime of this static is not clear to me, and I
> think this pattern works with at least some compilers, but I remember
> comments on previous patch review threads that this pattern isn't
> kosher.

I think this must be okay.  Some classic non-thread-safe C library 
functions essentially work that way, e.g.,

char *strerror(int errnum)
{
     static char buf[...];
     strcpy(buf, ....);
     return buf;
}

and then you can use the return pointer wherever you want.