https://gcc.gnu.org/bugzilla/show_bug.cgi?id=93812
Martin Sebor changed:
What|Removed |Added
Status|NEW |ASSIGNED
Assignee|unassigned at gcc dot gnu.org |msebor at gcc dot
gnu.org
Summary|[9/10 Regression] ICE in|[9/10 Regression] ICE on
|get_constant, at|redeclaration of an
|c-family/c-format.c:325 |attribute format function
||without protoype
--- Comment #4 from Martin Sebor ---
I'm not sure it makes sense to accept attribute format on functions without a
prototype. It certainly doesn't make sense to keep it after a declaration to
which it can't be applied has been seen. The only other kind of a declaration
that the attribute can meaningfully be applied is a variadic one but such is
not accepted after one without a prototype.
GCC 8 does accept the attribute on functions without a prototype and even seems
to do behave sensibly for "sensible" calls with the format argument in the
right place, but without -Wformat-nonliteral it doesn't do anything useful if
the format argument is not of the right type (e.g., because it's misplaced)
such as in:
__attribute__((__format__(__printf__, 1, 2))) void foo ();
void bar (void)
{
foo (1, "%s", 2); // no warning
}
(As an aside, not diagnosing at least the call if not the declaration seems
like a bug because GCC doesn't accept attribute format on variadic functions
where the format doesn't have a declared type compatible with char* and where
the first-to-check argument isn't the position of the ellipsis (i.e., it can't
be an arbitrary position within the variable argument list).)
A simple solution is to simply drop the attribute from functions without a
prototype with a warning like Clang does:
warning: '__format__' attribute only applies to Objective-C
methods, blocks, and non-K functions [-Wignored-attributes]
A slightly more involved but less intrusive (to user code) solution is to drop
it when a redeclaration is seen with a prototype.
