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

[John McCall]

On Oct 11, 2011, at 3:49 PM, Rafael Ávila de Espíndola wrote:
>> 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.
>> 
>> 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.
> 
> OK, so I think the consensus is
> 
> * The hard coded list will be removed after 3.0 branches and dragonegg 
> producing this attribute too.
> * The attribute should be part of the type
> * We should try to avoid allowing unsafe conversions like
> 
> int (*fptr)(jmp_buf) = &setjmp;
> 
> but a returns_twice function pointer can point to a function with an 
> otherwise matching type.

Sorry;  re-reading myself, I can see how this is a plausible interpretation,
but this is actually not what I meant.  Let me rephrase.

I think that, divorced from any other considerations, returns_twice would
belong in the type system.    However, doing so would introduce
significant complexity for basically zero value — if nothing else, the
use restrictions on these functions make it extremely unlikely that they
will ever be called indirectly.  Therefore I think it should stay a decl
attribute, and we can consider whether to make it a type attribute in
the off-chance that we ever see a compelling use case.

Note that even the template-metaprogramming use case will work
correctly with decl attributes.

John.