[cfe-commits] [patch] Add the returns_twice attribute to know functions

[Douglas Gregor]

On Oct 11, 2011, at 10:33 AM, John McCall wrote:

> On Oct 11, 2011, at 10:02 AM, Jakob Stoklund Olesen wrote:
>> On Oct 11, 2011, at 8:30 AM, Douglas Gregor wrote:
>>> Alternatively, we could take returns_twice out of the type system entirely. Is there value in declaring a function pointer to a returns_twice function, and having that be a distinct type from a similar function pointer to any function? If not, then perhaps returns_twice should be just a declaration attribute, and not have any impact on the type system. 
>> 
>> The thought of calling vfork() through a function pointer is very scary indeed, but the calling function needs to know that the called function may return twice.
>> 
>> The returns_twice attribute affects code generation in the caller, not in the callee.
>> 
>> I think that means it needs to be a type attribute.
>> 
>> However, if we allow this:
>> 
>> int (*fptr)(jmp_buf) = &setjmp;
>> 
>> We lose the information anyway.
> 
> Note that there *is* a subtyping relationship here:  a normal function can be safely used as a returns_twice function (which simply never actually returns twice).  But that's useless.

Yes, that's useless. And doing this properly in the type system means forbidding the code above, which means that putting returns_twice into the type system breaks currently-working code… for what?

> I agree that abstractly this ought to be a type attribute, and I don't think it would be difficult to take advantage of the current 'noreturn conversion' hooks to make the (unsafe) implicit conversion work, but I'm not sure there's any real value to adding this complexity.


I see added risk and added incompatibility with GCC, but no added value. I'd prefer that this go in as a Decl-only attribute for now, and we can revisit making it a type attribute down the road if we see any benefits.

	- Doug