svn commit: r280636 - head/include

[Tijl Coosemans]

On Thu, 26 Mar 2015 17:37:53 +1100 (EST) Bruce Evans <brde at optusnet.com.au> wrote:
> On Wed, 25 Mar 2015, Pedro Giffuni wrote:
>> On 03/25/15 21:14, Bruce Evans wrote:
>>> On Wed, 25 Mar 2015, Pedro F. Giffuni wrote:
>>>> Log:
>>>>  Temporarily revert 280458.
>>>> 
>>>>  GCC is still carries an old version of cdefs.h which doesn't
>>>>  accept multiple parameters for the nonnull attribute.
>>>>  Since this issue probably affects many ports in the tree
>>>>  we will revert it for now until gcc gets fixed.
>>> 
>>> Note that sys/cdefs.h is supposed to work with any version of
>>> gcc back to gcc-1, and does mostly work back to at least gcc-2.95.
>>> The whole point of sys/cdefs.h is to provide compatibity macros
>>> for old and other non-default compilers.  Standard compilers don't
>>> even have __attribute__(()).  So no changes in future versions
>>> of gcc will fix the previous commit.
>> 
>> cdefs.h still works for all versions of gcc back to gcc-1 AFAICT.
> 
> I now remember other bugs in it.  I think you put the varargs stuff
> in the non-gcc version.  That won't work compilers that don't support
> varargs for macros.  Neither will not changing the non-gcc version.
> 
> glibc (2.6 at least) avoids using varargs in its __nonnull() macro
> by using the same portable method that is used in many optional
> debugging statements including FreeBSD's KASSERT().  (KASSERT() is
> broken as designed.  It never needed this since it wasn't implmented
> until several years after C99 standardized  varargs for macros.)
> The macro takes a single arg consisting of a normal list of args
> enclosed in parentheses.  The extra parentheses are not passed to
> the __attribute__() list.  All invocations of the macro must be
> ugly to supply the parantheses.  The parentheses give a large
> syntactic difference, so the ugliness cannot be fixed easily by
> switching to varargs macros.  For KASSERT(), there would be about
> 7500 in /usr/src lines to clean up.  For __nonnull(), there would
> be only about lines 160 in /usr/src to change.  Mostly
> __nonnull(1) -> __nonnull((1)).  But __nonnull() is more likely to
> be (mis)used in ports.

Maybe introduce a __nonnull_all macro and leave __nonnull varargs-free:

#define __nonnull(x)	__attribute__((__nonnull__(x)))
#define __nonnull_all	__attribute__((__nonnull__))

Then in the rare cases where multiple arguments must be nonnull but
__nonnull_all doesn't apply you can use multiple __nonnull: 

int f(void *, void *, void *) __nonnull(1) __nonnull(2);

>> The reason why I had to revert the change is actually a systematic
>> bug in gcc: during it's build process gcc generates a new cdefs.h
>> from our headers. Attempting to use an older gcc from ports
>> that was build with the broken mono-parameter __nonnull() ended
>> up causing breakage in any code using signal.h or pthreads.h.
> 
> I see.  gcc's "fixed" headers cause lots of problems.

I've complained about this multiple times in the past.  The gcc ports
should not install these "fixed" headers.

Pedro, by reverting this commit you only allow this problem to persist,
so please reapply it.  You also shouldn't wait weeks before applying
the next commit.  No amount of waiting is enough.  There will always be
users bitten by it.  The problem is in the ports.  It needs to be fixed
there.  If you receive any problem reports that are caused by this gcc
problem, forward them to the gcc port maintainer.