svn commit: r268137 - head/sys/sys

[Bruce Evans]

On Sat, 20 Jun 2015, Pedro Giffuni wrote:

> On 06/19/15 12:23, Bruce Evans wrote:
>> On Fri, 19 Jun 2015, Dimitry Andric wrote:
>> 
>>> On 19 Jun 2015, at 17:02, Pedro Giffuni <pfg at freebsd.org> wrote:
>>>> 
>>>>> On 19/06/2015 05:16 a.m., David Chisnall wrote:
>>>>>> I only just caught this (having seen the fallout from NetBSD doing the 
>>>>>> same thing in a shipping release and the pain that it’s caused):
>>>>>> __weak is a reserved keyword in Objective-C, please pick another name 
>>>>>> for this.  This in cdefs.h makes it impossible to include any FreeBSD 
>>>>>> standard headers in Objective-C programs (of which we have a couple of 
>>>>>> hundred in ports) if they use any of the modern Objective-C language 
>>>>>> modes.
>>> ...
>>>> Closely related to this, we are redefining _Noreturn, which is a reserved 
>>>> keyword in C11.
>>> 
>>> No, sys/cdefs.h has:
>>> 
>>>   254  /*
>>>   255   * Keywords added in C11.
>>>   256   */
>>>   257
>>>   258  #if !defined(__STDC_VERSION__) || __STDC_VERSION__ < 201112L || 
>>> defined(lint)
>>> [...]
>>>   284  #if defined(__cplusplus) && __cplusplus >= 201103L
>>>   285  #define _Noreturn               [[noreturn]]
>>>   286  #else
>>>   287  #define _Noreturn               __dead2
>>>   288  #endif
>>> [...]
>>>   320  #endif /* __STDC_VERSION__ || __STDC_VERSION__ < 201112L */
>>> 
>>> So the whole block redefining all the _Xxx identifiers is skipped for
>>> C11 and higher.
>> 
>> I probably pointed this out incorrectly to Pedro.
>> 
>> All uses of _Noreturn are still broken, and also ugly.  __dead2 is the
>> gcc-2 compatible version of the gcc-1 compatible macro __dead.  It is
>> syntactically different from __dead and _Noreturn.  It must be placed
>> after the function parameter list instead of in the function type
>> declarator because old versions of gcc only accept attributes there.
>> __dead and presumably _Noreturn must be placed in the function type
>> declarator.  This is incompatible, and also uglier.
>> 
> I was thinking that _Noreturn can be fixed for older compilers
> less disruptively.
>
> I haven't tested the attached patch the idea is to resurrect
> __dead and use it for _Noreturn.
>
>> Correct version with ugly declarations:
>> 
>> __dead void
>>     foo(void) __dead2;
>> 
>
> With the patch we would use:
>
> __Noreturn void
>   foo(void) _dead2;
>
> Which is still ugly but C11-ish.

That asks for the same problems as defining __weak.

Why not just don't use _Noreturn?  It is an unimprovement on the gcc
attribute.  The attribute works at the beginning or end, while Noreturn
only works at the end.  I checked this:

pts/23:bde at freefall:~/s> cc -O -S a.c -std=c11
a.c:3:26: error: '_Noreturn' keyword must precede function declarator
_Noreturn void bar(void) _Noreturn;
                          ^

I also checked if __dead2 works at the beginning in gcc-2.95.4.  It
does, but other headers are broken, so the support for gcc-2.95.4 in
sys/cdefs.h is almost useless:

pts/23:bde at freefall:~/s> gcc295 -O -S a.c
In file included from /usr/include/machine/_types.h:6,
                  from /usr/include/sys/_types.h:33,
                  from /usr/include/x86/endian.h:37,
                  from /usr/include/machine/endian.h:6,
                  from /usr/include/sys/types.h:44,
                  from a.c:1:
/usr/include/x86/_types.h:161: syntax error before `__gnuc_va_list'
/usr/include/x86/_types.h:161: warning: data definition has no type or storage c
lass

Here is the broken part of x86/_types.h:

%%%
#ifdef __GNUCLIKE_BUILTIN_VARARGS
typedef __builtin_va_list	__va_list;	/* internally known to gcc */
#elif defined(lint)
typedef char *			__va_list;	/* pretend */
#endif
#if defined(__GNUC_VA_LIST_COMPATIBILITY) && !defined(__GNUC_VA_LIST) \
     && !defined(__NO_GNUC_VA_LIST)
#define __GNUC_VA_LIST
meta: next is line 161:
typedef __va_list		__gnuc_va_list; /* compatibility w/GNU headers*/
#endif
%%%

Despite (rather, because of) mounds of ifdefs to support old gcc, it doesn't
actually work.

<stdarg.h> is honestly broken.  It knows that it depends on the gcc feature
__GNUCLIKE_BUILTIN_STDARG, and aborts with an #error when it is not available.
All of the _GNUCLIKE_BUILTIN_* macros for variadic args are controlled by
sys/cdefs.h.  They are defined when the compiler is either gcc newer than
gcc-2.95 or is icc.

Bruce