Re: [PATCH v4 3/8] bits: introduce fixed-type genmasks

[Vincent Mailhol]

On 06/03/2025 at 04:45, Andy Shevchenko wrote:
>>> But GENMASK_U128() becomes a special case now.
>>> The 128-bit GENMASK is unsued, but it's exported in uapi. Is there any
>>> simple way to end up with a common implementation for all fixed-type
>>> GENMASKs?
>>
>> What bothers me is that the 128 bit types are not something available on
>> all architectures, c.f. the CONFIG_ARCH_SUPPORTS_INT128. So, I would
>> need a U128() equivalent to the ULL() but which does not break on
>> architectures which do not support 128 bits integers.
>>
>> This is where I am stuck. If someone can guide me on how to write a
>> robust U128() macro, then I think the common implementation could be
>> feasible.
> 
> I think we may leave that U128 stuff alone for now.

I found the solution! The trick is to use type_max() from overflow.h.

With this, GENMASK_TYPE() becomes:

  #define GENMASK_TYPE(t, h, l)					\
  	((t)(GENMASK_INPUT_CHECK(h, l) +			\
  	     (type_max(t) << (l) &				\
  	      type_max(t) >> (BITS_PER_TYPE(t) - 1 - (h)))))

and works with all the GENMASK variants, including the U128 one! The
unit tests under lib/test_bits.c are all green.

Of course, this does *not* work in assembly. But as explained before,
GENMASK_TYPE() is guarded by a #if !defined(__ASSEMBLY__), so all good!

The question raised by Yury on whether or not we should keep
__GENMASK_U128() in the uapi still remains. And in full honesty, I will
not touch that one. This is not in the scope of this series.


Yours sincerely,
Vincent Mailhol