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

[Rikard Falkeborn]

Hi Masahiro,

On Wed, Aug 07, 2019 at 12:19:36AM +0900, Masahiro Yamada wrote:
> Hi Rikard,
> 
> 
> On Tue, Aug 6, 2019 at 4:55 AM Rikard Falkeborn
> <rikard.falkeborn@xxxxxxxxx> wrote:
> >
> > On Sun, Aug 04, 2019 at 03:45:16PM +0900, Masahiro Yamada wrote:
> > > On Sun, Aug 4, 2019 at 3:36 AM Rikard Falkeborn
> > > <rikard.falkeborn@xxxxxxxxx> wrote:
> > > >
> > > > On Sat, Aug 03, 2019 at 12:12:46PM +0900, Masahiro Yamada wrote:
> > > > > On Sat, Aug 3, 2019 at 12:03 PM Masahiro Yamada
> > > > > <yamada.masahiro@xxxxxxxxxxxxx> wrote:
> > > > >
> > > > > >
> > > > > > BTW, v2 is already inconsistent.
> > > > > > If you wanted GENMASK_INPUT_CHECK() to return 'unsigned long',,
> > > > > > you would have to cast (low) > (high) as well:
> > > > > >
> > > > > >                (unsigned long)((low) > (high)), UL(0))))
> > > > > >
> > > > > > This is totally redundant, and weird.
> > > > >
> > > > > I take back this comment.
> > > > > You added (unsigned long) to the beginning of this macro.
> > > > > So, the type is consistent, but I believe all casts should be removed.
> > > >
> > > > Maybe you're right. BUILD_BUG_ON_ZERO returns size_t regardless of
> > > > inputs. I was worried that on some platform, size_t would be larger than
> > > > unsigned long (as far as I could see, the standard does not give any
> > > > guarantees), and thus all of a sudden GENMASK would be 8 bytes instead
> > > > of 4, but perhaps that is not a problem?
> > >
> > >
> > > How about adding (int) cast to BUILD_BUG_ON_ZERO() ?
> >
> > I'll have a look.
> 
> 
> 
> I found a more important problem in this patch.
> 
> You used __is_constexpr(), which is defined in <linux/kernel.h>.
> 
> This header does not include <linux/kernel.h>,
> so this header is not self-contained anymore.
> 
> The following test code fails to build:
> 
> #include <linux/bits.h>
> unsigned long foo(unsigned long in_bits)
> {
>         return in_bits & GENMASK(5, 3);
> }
> 
> 
> 
> However, you cannot include <linux/kernel.h> from <linux/bits.h>.
> See the log of 8bd9cb51daac89337295b6f037b0486911e1b408
> 
> This header was split out to not pull in <linux/bitops.h>
> Including <linux/kernel.h> pulls in <linux/bitops.h> again.
> 
> 
> In summary, please use __builtin_constant_p()
> instead of __is_constexpr().
> 
> 
> You can shorten __builtin_constant_p(high) && __builtin_constant_p(low)
> into __builtin_constant_p((low) > (high)).
> 
> 
> How about this?
> 
> 
> #define GENMASK_INPUT_CHECK(high, low) \
>        BUILD_BUG_ON_ZERO(__builtin_choose_expr( \
>               __builtin_constant_p((low) > (high)), (low) > (high), 0))
> 
> 
> -- 
> Best Regards
> Masahiro Yamada

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.

Best Regards,
Rikard Falkeborn