[cfe-commits] r58008 - /cfe/trunk/lib/Sema/SemaDecl.cpp

[Douglas Gregor]

On Oct 22, 2008, at 6:39 PM, Argiris Kirtzidis wrote:

> Douglas Gregor wrote:
>>
>> I don't think this is the right fix. If we find an invalid  
>> redeclaration, we should just drop the redeclaration.
>
> Makes sense, but before I apply some fix I'd like to make sure, is  
> this definitely the intended behavior ? Currently there's a test  
> that assumes that name lookup finds the latest redeclaration, even  
> if it's invalid.
> From test/Sema/predefined-function.c :
>
> int eli(float b); // expected-error {{previous declaration is here}}
> int b(int c) {return 1;}
>
> int foo();
> int foo()
> {
>   int eli(int (int)); // expected-error {{conflicting types for  
> 'eli'}}
>   eli(b);  #1
>   return 0;   }
>
> At #1, it's supposed to pick up the invalid redeclaration at the  
> previous line.

I actually don't know the intent, but I can ask around. I could  
imagine, for example, that one might actually want to do this: if we  
get an invalid redeclaration, have PushOnScopeChains replace the  
existing declaration with the new declaration, so that type-checking  
can continue without having multiple, inconsistent declarations of the  
same thing in scope.

But, that's a bit of work when we're not sure what the intent was, so  
I'm going to withdraw my objection until such time as the thing I  
describe below comes to pass, or we decide that we don't really need  
overloading in C :)

>
>> There may still be uses for overloading when we're not in C++ mode  
>> (e.g., for <tgmath.h>).
>
> Could you explain a bit more about this ?

<tgmath.h> is a C99 header that has several "type-generic" macros in  
it, which are versions of common C standard library functions (acos,  
atan, exp, etc.) whose type is the same as their input type, e.g.,  
acos(float) is a float, acos(double) is a double, etc. Now, you can't  
actually implement <tgmath.h> in C99, so one tends to use some kind of  
builtins. However, one possible implementation is to have a little  
incantation that allows C++ overloading in C, and to implement  
tgmath.h using C++ overloading. I'd like to keep that possibility open.

	- Doug