Michal, On Sep 12, 2014, at 4:37 PM, Michal Nazarewicz <min...@mina86.com> wrote:
> On Fri, Sep 12 2014, Andrew Morton <a...@linux-foundation.org> wrote:
>> On Thu, 11 Sep 2014 23:39:36 +0200 Michal Nazarewicz <min...@mina86.com>
>> wrote:
>>
>>> Because min and max macros use the same variable names no matter
>>> how many times they are called (or how deep the nesting of their
>>> calls), each time min or max calls are nested, the same variables
>>> are declared. This is especially noisy after min3 and max3 have
>>> been changed to nest min/max calls.
>>>
>>> Using __COUNTER__ solves the problem since each variable will get
>>> a unique number aadded to it. The code will still work even if
>>> the compiler does not support __COUNTER__, but then the protection
>>> from shadow warning won't work.
>>>
>>> The same applies to min_t and max_t macros.
>>>
>>> ...
>>>
>>> --- a/include/linux/kernel.h
>>> +++ b/include/linux/kernel.h
>>> @@ -695,15 +695,27 @@ static inline void ftrace_dump(enum ftrace_dump_mode
>>> oops_dump_mode) { }
>>> #endif /* CONFIG_TRACING */
>>>
>>> /*
>>> + * Preprocessor magic generating unique identifiers to avoid -Wshadow
>>> warnings
>>> + * used by min, max, min_t and max_t macros. cnt is __COUNTER__, op is the
>>> + * comparison operator; tx (ty) is type of the first (second) argument,
>>> + * xx (yy) is name of a temporary variable to hold the first (second)
>>> argument,
>>> + * and x (y) is the first (second) argument.
>>> + */
>>> +#define _min_max_var(cnt, base) _mm_ ## cnt ## base
>>> +#define _min_max__(op, tx, xx, x, ty, yy, y) ({ \
>>> + tx xx = (x); \
>>> + ty yy = (y); \
>>> + (void) (&xx == &yy); \
>>> + xx op yy ? xx : yy; })
>>> +#define _min_max_(cnt, op, tx, x, ty, y) \
>>> + _min_max__(op, tx, _min_max_var(cnt, a), x, ty, _min_max_var(cnt, b), y)
>>> +#define _min_max(...) _min_max_(__COUNTER__, __VA_ARGS__)
>>
>> The fact that __COUNTER__ is used in compiler-gcc4.h but not in
>> compiler-gcc3.h makes me suspicious about its availability?
>
> http://gcc.gnu.org/ml/gcc-patches/2007-05/msg01579.html so looks like it
> has 7 years. But as the commit message says, the code will still work,
> even w/o working __COUNTER__.
>
>> I do think that [1/2] made the code significantly worse-looking
>
> Oh? I actually thought [1/2] makes it nicer by having a single place
> where the min/max logic is implemented.
It does have that going for it.
>> and this one is getting crazy. How useful is W=2 anyway?
>
> I actually do agree with that. I didn't have high hopes about getting
> this patch accepted, but wanted to send it out to show that it can be
> done, if it's really deemed useful.
Well, I learned something from it. Thank you for teaching this old dog a new
trick.
>> Has anyone found a bug using it? The number of warnings in default
>> builds is already way too high :(
--
Mark Rustad, Networking Division, Intel Corporation
signature.asc
Description: Message signed with OpenPGP using GPGMail
