RFC: Warnings silenced when macros from system headers are used (PR c/78000, c/71613)

[Joseph Myers]

On Wed, 2 Nov 2016, Jason Merrill wrote:

> It seems to me that the general principle is that we should consider
> the location where the thing we're warning about is happening.  In
> 
>    float_var = LLONG_MIN;
> 
> The relevant location is that of the assignment, not the constant on
> the RHS.  In your ?: example, a simple answer would be to warn based

I'm not sure we track locations well enough to handle that yet.

Say you have an implicit conversion of a function argument to the type 
from the prototype and something about that conversion should be warned 
about.  Then you should obviously warn if the argument is a macro from a 
system header but the call is outside a system header.  But say the 
function call itself comes from a macro defined in a system header - you 
should still warn if the user passed an argument of the wrong type, even 
if that argument is a macro from a system header.

That is:

/* System header.  */
int __foo (int);
/* This sort of macro to avoid accidental interposition issues has been 
   discussed lately on libc-alpha, so it's a realistic example.  */
#define foo(x) (0 + __foo (x))
/* User code.  */
v = foo (NULL);

should warn because the call to __foo logically results from user code 
even though both NULL and foo are macros defined in system headers.  I'm 
not sure what the locations look like here.

-- 
Joseph S. Myers
joseph@codesourcery.com