On Wed, Sep 06, 2023 at 03:08:33PM -0700, Mike Fulton wrote:
> Hello,
>
> The m4 dev line ( git://git.savannah.gnu.org/m4.git )
> has the following code:
>
> ```
> #if 4 < __GNUC__ + (6 <= __GNUC_MINOR__)
> # pragma GCC diagnostic push
> # pragma GCC diagnostic ignored "-Wformat-nonliteral"
> #endif
> ```
That _expression_ appears to be designed to reject gcc 3.5 and older,
while accepting 3.6 and newer.
Right - that's my understanding.
I also wonder if modern gnulib provides a better way to test things
than just open-coding it (as Bruno has done a lot of work testing
features of newer compilers). This may be one case where ignoring the
diagnostic for older builds is desirable - there comes a point where
you can't please both old and new compilers simultaneously, but the
target that should be warning-free is the newer compiler.
Great question - this seems to be something added before clang and I
wonder if there is a better way to manage diagnostics that are meant
to be 'sort of common?' between GCC and clang. I've just started
using clang on z/OS so I'm not sure how this mapping is done for
GCC compatibility.
That is, maybe it's time to just rewrite the check to something as
simple as #if 8 <= __GNUC__ (that is, only add the pragmas on a much
newer, albeit arbitrary, point in time)?
Conversely, I see that m4 uses gnulib's lib/c-stack.c, which also has:
/* Pacify GCC 9.3.1, which otherwise would complain about segv_handler. */
# if 4 < __GNUC__ + (6 <= __GNUC_MINOR__)
# pragma GCC diagnostic ignored "-Wsuggest-attribute=pure"
# endif
so it's not like m4 is doing anything radically different from gnulib
when it comes to pragmas for ignoring certain diagnostics; but thd
difference may be which diagnostic is being ignored.
>
> This produces an error on the clang compiler I am using on z/OS, which is
> at 4.2 with the xasprintf code that follows, e.g.
>
> ```
> str = xasprintf (fstart, width, ARG_INT(argc, argv));
> ```
>
> If I change the check to:
>
> ```
> #if 4 < __GNUC__ + (2 <= __GNUC_MINOR__)
> # pragma GCC diagnostic push
> # pragma GCC diagnostic ignored "-Wformat-nonliteral"
> #endif
> ```
>
> all is good - it compiles clean. The question is whether the check for 6 is
> too high or if the clang compiler on z/OS has a bug (or something else)?
What does clang claim to be? That is, what do you get for
print '__GNUC__ __GNUC_MINOR__' | clang -E - | tail -n1
On both z/OS as well as my Mac I see:
4 2
and what is the actual compiler error you got when the build failed on
the original source?
CC format.o
format.c:360:28: error: format string is not a string literal [-Werror,-Wformat-nonliteral]
str = xasprintf (fstart, width, ARG_INT(argc, argv)); ^~~~~~
> I haven't been able to determine what level of gcc provides this diagnostic.
What if we just acknowledged the clang compiler and changed the check to:
#if 4 < __GNUC__ + (6 <= __GNUC_MINOR__) || defined(__clang__)
The current clang compiler (not an old one) states that it's __GNU__C is 4 and
it's __GNUC_MINOR__ is 2 (for both the Mac and z/OS - and I assume other
platforms too - but I don't have access to other systems). So - I would like to
propose that we extend the macro as I described above to add a check for
the clang compiler in addition to the GNUC compiler to enable the #pragmas
which will eliminate warnings (or in my case errors because I flag warnings as
errors with my build). The fix is:
if 4 < __GNUC__ + (6 <= __GNUC_MINOR__) || defined(__clang__)
I would be happy to either provide a patch or have someone with
the right authority, provide the patch.
?
>
> thanks, mike
--
Eric Blake, Principal Software Engineer
Red Hat, Inc.
Virtualization: qemu.org | libguestfs.org