Re: [PATCH] attribute: const/pure defaults to unsequenced/reproducible

[Paul Eggert]

On 2024-06-05 08:32, Basil L. Contovounesios wrote:

> > It's also too bad that N2956 has this typo, and that it doesn't mention
> > GCC's returns-exactly-once constraints on pure and const functions.
>
> Do you have a pointer to where this constraint is documented?  I'm not
> sure what it means, and I can't find it in my copy of the gcc-13 manual.

It's not stated directly, but it's there. It should be documented 
better. I've confirmed it by compiling programs and looking at the code, 
and via messages from GCC implementers.

For example, compile this with 'gcc -O2 -S':

  int difftime (long, long) __attribute__ ((pure));

  int
  f ()
  {
    return difftime (42, 27) - difftime (42, 27);
  }

and you'll see that the calls to difftime are optimized away and f 
always returns 0. This optimization would be invalid if difftime did not 
return, or returned multiple times.

The C23 spec for [[unsequenced]] and [[reproducible]] does not mention 
this issue, which is a significant oversight.