On Tue, May 30, 2023 at 11:23:42AM +0200, Peter Zijlstra wrote:
> Yes, it's a wee bit more involved, but I'm thinking it gives a fair
> amount of flexibility and we don't need to ret rid of
> -Wdeclaration-after-statement.
So I made all that work and .. Yes, you're absolutely right.
Busting -Wdeclaration-after-statement is the right thing to do for
guards.
So then I came up with:
#define __ptr_guard(_guard, _name) \
guard_##_guard##_t _name __cleanup(guard_##_guard##_cleanup)
#define ptr_guard(_guard, _name) \
__diag(push) \
__diag(ignored "-Wdeclaration-after-statement") \
__ptr_guard(_guard, _name) \
__diag(pop)
#define guard_init(_guard, _var...) \
guard_##_guard##_init(_var)
#define named_guard(_guard, _name, _var...) \
ptr_guard(_guard, _name) = guard_init(_guard, _var)
#define guard(_guard, _var...) \
named_guard(_guard, __UNIQUE_ID(guard), _var)
#define scoped_guard(_guard, _var...) \
for (__ptr_guard(_guard, scope) = guard_init(_guard, _var), \
*done = NULL; !done; done = (void *)1)
And that all (mostly) works on clang, but not GCC :-( GCC refuses to
accept _Pragma() inside an expression.
So I now have that ptr_guard() with push/pop for clang but without for
GCC, which means that only clang has a fighting chance to report
-Wdeclaration-after-statement warns until such a time as that we can get
GCC 'fixed'.
https://godbolt.org/z/5MPeq5W6K
FWIW: the work-in-progress patches I have are here:
https://git.kernel.org/pub/scm/linux/kernel/git/peterz/queue.git/log/?h=core/guards
