[Bug c++/68052] No printf format warnings in C++ code for macro in system headers

["gcc at blino dot org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68052

--- Comment #3 from Olivier Blin <gcc at blino dot org> ---
Quoting Martin Sebor's answer from the gcc-help mailing list:
https://gcc.gnu.org/ml/gcc-help/2015-10/msg00128.html

> I'm inclined to say it's a bug. I recommend opening it in Bugzilla.
> The only way I was able to work around it is by changing MY_PRINTF
> from a function-like macro to an object macro like so:
> 
>   #define MY_PRINTF printf
> 
> The problem isn't specific to -Wformat but can be reproduced with
> other warnings as well. For example, replacing the contents of
> the header in the test case with this:
> 
>   #pragma GCC system_header
> 
>   static inline void __attribute__ ((deprecated)) foobar (void) { }
> 
>   #define foo foobar
>   #define bar() foobar()
> 
> and the body of main with
> 
>   foo ();
>   bar ();
> 
> emits a warning for each of foo() and bar() in C but just one,
> for foo(), in C++.
> 
> While stepping through the code that checks to see whether or not
> a diagnostic should be issued (diagnostic_issue_diagnostic) I see
> that the difference in the bar() case is in the return value of
> the linemap_macro_expansion_map_p() function which is documented
> to
> 
>   Return TRUE if MAP encodes locations coming from a macro
>   replacement-list at macro expansion point.
> 
> and implemented by testing the location MAP's REASON member for
> equality to the LC_ENTER_MACRO constant. In C, the REASON member
> is set to LC_LEAVE while in C++ to LC_ENTER_MACRO. In C, the
> call to bar is then treated as an ordinary token whose location
> (i.e., bar's invocation) determines whether or not to issue the
> diagnostic. In C++ the reference to bar is recognized as a macro
> expansion and the function proceeds to find the token it expands
> to and use that to determine whether or not to diagnose it. And
> since the token bar() expands to is defined in a system header,
> C++ doesn't issue the warning.
> 
> In the case of foo(), the REASON is LC_LEAVE in both front-ends
> and so they both diagnose it.
> 
> Martin