Re: CVS commit: src/sys/sys

[Roland Illig <roland.illig%gmx.de@localhost>]

Am 05.04.2025 um 02:07 schrieb Taylor R Campbell:
>> Module Name:    src
>> Committed By:   rillig
>> Date:           Fri Apr  4 20:52:32 UTC 2025
>>
>> Modified Files:
>>         src/sys/sys: cdefs.h
>>
>> Log Message:
>> sys/cdefs.h: fix __predict_true and __predict_false for lint
>>
>> -#define	__predict_true(exp)	__builtin_expect((exp) ? 1 : 0, 1)
>> -#define	__predict_false(exp)	__builtin_expect((exp) ? 1 : 0, 0)
>> +#define	__predict_true(exp)	__builtin_expect(/*CONSTCOND*/(exp) ? 1 : 0, 1)
>> +#define	__predict_false(exp)	__builtin_expect(/*CONSTCOND*/(exp) ? 1 : 0, 0)
>
> This seems wrong to me.  Why should we tell lint that _every_ input to
> __predict_true/false is constant?
>
> What lint objected to in this case is that ALIGNED_POINTER(...) is
> always 1 on x86.  But it's not always 1 on all architectures.

Agreed.

Would it be narrow enough to add /*CONSTCOND*/ to the definition of
ALIGNED_POINTER on i386 and amd64? Or would you consider this still too
broad?

The next alternative would be to mark each call to ALIGNED_POINTER as
/*CONSTCOND*/, but that would be additional work in several places.

Roland