On 12/05/2021 09.48, Arnd Bergmann wrote:

On Tue, May 11, 2021 at 10:39 PM Rikard Falkeborn
[off-list ref] wrote:

quoted

On Tue, May 11, 2021 at 08:53:53PM +0900, Tetsuo Handa wrote:

quoted

quoted

#define GENMASK_INPUT_CHECK(h, l) \
     (BUILD_BUG_ON_ZERO(__builtin_choose_expr( \
          __builtin_constant_p((l) > (h)), (l) > (h), 0)))

__GENMASK() does not need "h" and "l" being a constant.

Yes, small_const_nbits(size) in find_next_bit() can guarantee that "size" is a
constant and hence "h" argument in GENMASK_INPUT_CHECK() call is also a constant.
But nothing can guarantee that "offset" is a constant, and hence nothing can
guarantee that "l" argument in GENMASK_INPUT_CHECK() call is also a constant.

Then, how can (l) > (h) in __builtin_constant_p((l) > (h)) be evaluated at build time
if either l or h (i.e. "offset" and "size - 1" in find_next_bit()) lacks a guarantee of
being a constant?

So the idea is that if (l > h) is constant, __builtin_constant_p should
evaluate that, and if it is not it should use zero instead as input to
__builtin_chose_expr(). This works with non-const inputs in many other
places in the kernel, but apparently in this case with a certain
compiler, it doesn't so I guess we need to work around it.

I have a vague memory that __builtin_constant_p() inside of
__builtin_choose_expr()
always evaluates to false because of the order in which the compiler processes
those: If constant-folding only happens after __builtin_choose_expr(), then
__builtin_constant_p() has to be false.

It's more complicated than that. __builtin_constant_p on something which
is a bona-fide Integer Constant Expression (ICE) gets folded early to a
1. And then it turns out that such a __builtin_constant_p() that folds
early to a 1 can be "stronger" than a literal 1, in the sense that when
used as the controlling expression of a ?: with nonsense in the false
branch, the former is OK but the latter fails:

https://lore.kernel.org/lkml/c68a0f46-346c-70a0-a9b8-31747888f05f@rasmusvillemoes.dk/ (local)

Now what happens when the argument to __builtin_constant_p is not an ICE
is a lot more complicated. The argument _may_ be so obviously
non-constant that it can be folded early to a 0, hence still be suitable
as first argument to __b_c_e. But it is also possible that the compiler
leaves it unevaluated, in the "hope" that a later optimization stage
could prove the argument constant. And that's the case where __b_c_e
will then break, because that can't be left unevaluated for very long -
the very _type_ of the result depends on which branch is chosen.

tl;dr: there's no "order in which the compiler processes those", __b_c_p
can get evaluated (folded) early, before __b_c_e inspects it, or be left
for later stages.

Rasmus