Re: [RFC PATCH] compiler.h: give up __compiletime_assert_fallback()

From: Masahiro Yamada
Date: Sat Aug 25 2018 - 14:12:46 EST


Hi Daniel,


2018-08-21 17:11 GMT+09:00 Daniel Santos <daniel.santos@xxxxxxxxx>:
> On 08/19/2018 03:25 PM, Nick Desaulniers wrote:
>> + gbiv who wrote this cool paste (showing alternatives to
>> _Static_assert, which is supported by both compilers in -std=gnu89,
>> but not until gcc 4.6): https://godbolt.org/g/DuLsxu
>>
>> I can't help but think that BUILD_BUG_ON_MSG should use
>> _Static_assert, then have fallbacks for gcc < 4.6.
>
> Unfortunately _Static_assert is a woefully inadequate replacement
> because it requires a C constant expression. Example:
>
> int a = 1;
> _Static_assert(a == 1, "a != 1");
>
> results in "error: expression in static assertion is not constant."


You are right.




I tried diagnose_if from Clang:

static inline void assert_d(int i) __attribute__((diagnose_if(!i, "oh
no", "error")))
{
}



Clang is silent about

int a = 1;
assert_d(a);




But,

if (0)
assert_d(0);


is error.

Hence, it cannot be used for BUILD_BUG().







Anyway, I will just try to rebase this patch and send it to Linus.



> Language standards tend to shy away from defining implementation details
> like optimizations, but we need to have completed a good data flow
> analysis and constant propagation in order to do BUILD_BUG_ON_MSG, et.
> al.; this is why they only work when optimizations are enabled. As the
> optimizer improves, new expressions can be used with BUILD_BUG_ON*. I
> did an analysis of this back in 2012 of how various types of variables
> could be resolved to constants at compile-time and how that evolved from
> gcc 3.4 to 4.7:
>
> https://drive.google.com/open?id=1cQRAAOzjFy6Aw7CDc4QauHvd_spVkd5a
>
> This changed again when -findirect-inline was added -- i.e.,
> BUILD_BUG_ON could be used on parameters of inline functions even when
> called by pointer, although the caller needed __flatten in some cases --
> a bit messy.
>
> Daniel



--
Best Regards
Masahiro Yamada