Re: [PATCH] linux/bits.h: Add compile time sanity check of GENMASK inputs

[Rikard Falkeborn <rikard.falkeborn@gmail.com>]

On Tue, Aug 06, 2019 at 02:15:54PM -0700, Joe Perches wrote:
> On Tue, 2019-08-06 at 21:27 +0200, Rikard Falkeborn wrote:
> > On Wed, Aug 07, 2019 at 12:19:36AM +0900, Masahiro Yamada wrote:
> > > How about this?
> > > #define GENMASK_INPUT_CHECK(high, low) \
> > >        BUILD_BUG_ON_ZERO(__builtin_choose_expr( \
> > >               __builtin_constant_p((low) > (high)), (low) > (high), 0))
> > Thanks for the feedback, your version looks much cleaner than mine. I
> > *think* I had a reason for using __is_constexpr() instead of
> > __builtin_constant_p but I'll try a full rebuild to see if something
> > comes up.
> 
> Perhaps a statement expression so high and low aren't possibly
> evaluated multiple times?
> 
> #define GENMASK_INPUT_CHECK(high, low)				\
> ({								\
> 	typeof(high) _high = high;				\
> 	typeof(low) _low = low;					\
> 	BUILD_BUG_ON_ZERO(__builtin_constant_p(_low > _high,	\
> 					       _low > _high,	\
> 					       0))		\
> })
> 
> 

That doesn't work I think (even after adding __builtin_choose_expr).
Even so, high and low are not evaluated multiple times (they're not
evaluated at all in GENMASK_INPUT_CHECK, if they were, the arguments
would be evaluated twice since they're evaluated in __GENMASK as well.

__builtin_constant_p does not seem to evaluate it's expression (even
though I didn't manage to find that spelled out in the docs, but since
__builtin_constant_p is evaluated at compile time it makes sense that it
doesn't), and __builtin_choose_expr does not evaluate the operand that
is not chosen (this is actually in the docs). Even if it was,
BUIlD_BUG_ON_ZERO uses sizeof its argument, which is evaluated at compile
time (unless the argument is a VLA).

So this should be safe.