Re: [1003.1(2008)/Issue 7 0000074]: Pointer Types Problem

[Don Cragun <dcragun@xxxxxxxxx>]

Wojtek, Fred, David, and Douglas,
I apologize for my confusion here.  I have had discussions with some  
of you concerning these points over the last decade, but I lost access  
to the logs of some of those discussions when Sun went through a work- 
force reduction earlier this year.

I believe the open points are the following:
1.  Was the resolution of Austin Group Interpretation #003:
        http://www.opengroup.org/austin/interps/uploads/40/1068/ai-003.txt
    correct when it was written?  In particular was the statement:
        fptr = (int (*)(int))dlsym(handle, "my_function");
    required to generate a diagnostic by the C99 Standard as of
    July 9, 2002?  And, was the statement:
        *(void **)(&fptr) = dlsym(handle, "my_function");
    a functionally equivalent substitute that did not require a
    diagnostic?  In both cases, I'm using "diagnostic" in the sense
    of the C standard; not the POSIX standard.
2.  Does the current C standard (i.e., the approved C99 standard
    including approved TCs; but not including drafts of upcoming
    changes) require a compiler to generate a diagnostic for either
    of the statements:
        fptr = (int (*)(int))dlsym(handle, "my_function");
        *(void **)(&fptr) = dlsym(handle, "my_function");
    today?

Thanks,
Don

On Aug 29, 2009, at 2:34 PM, Wojtek Lerch wrote:

> Fred J. Tydeman wrote:
> > On Sat, 29 Aug 2009 10:22:34 +0100, Geoff Clare wrote:
> ...
> > > Personally, I can't find anything in C99 that requires the warning,
> > > but just because I can't find it doesn't mean it's not there.
> > I am using WG14 N1362 (C1x) as my reference document, as it is C99
> > plus all the Technical Corrigendums plus some of the approved changes
> > for the next version of C.
> >
> > A conversion of a pointer can be done via 6.5.4 (Cast operators) or
> > 6.5.16.1 (Simple assignment).  Neither of those two has an explicit
> > constraint on pointer to void being converted to pointer to function.
> There is an important difference: 6.5.16.1 has a long list of  
> combinations of types that are allowed in a simple assignment, and  
> none of them mixes a function pointer with a data pointer, whereas  
> the constraints of 6.5.4 allow casts from any scalar type to any  
> other scalar type, including all possible pointer types.  This means  
> that converting between a function pointer and a data pointer by  
> simple assignment violates the constraints of 6.5.16.1 and therefore  
> requires a diagnostic, while a similar conversion by an explicit  
> cast is not a constraint violation, and does not require a  
> diagnostic, even when it has undefined behaviour.
>
> > 6.3.2.3 (Pointers) makes no mention of pointer to void being  
> > converted
> > to pointer to function, therefore, that conversion is implicitly
> > undefined behaviour (by 4. Conformance, paragraph 2).
> Right.
>
> > Aside: the
> > constraints for these kinds of conversions are in 6.5.4 Cast
> > operators.
> Pedantically speaking, the constraints for the various kinds of  
> syntactic costructs that involve conversions (such as casts,  
> assignments, or binary operations on values that may require  
> conversions to a common type) are described separately for each of  
> those constructs.  The C standard does not specify any general  
> constraints that apply to all conversions.
>
> > 5.1.1.3 (Diagnostics) only requires a diagnostic for a violation of
> > syntax or constraints.  There is no required diagnostic for implicit
> > undefined behaviour.