libspl sys/sysmacros.h: Fix P2ROUNDUP_TYPED to not trigger integer overflow

[perfinion]

The original P2ROUNDUP_TYPED macro contains a -x which triggers PaX's
integer overflow detection for unsigned integers. Replace the macro with
an equivalent version that does not trigger the overflow.

Signed-off-by: Jason Zaman [email protected]

fixes #2505

[ryao]

After talking with @perfinion about this, I am confident that this is mathematically equivalent and can prove it. Here are axioms:

a. (-(x)) === (((x) - 1)) === ((x) + 1) under two's complement
b. (x & y) === (((x)) | (~(y))) under De Morgan's law
c. ~(~x) === x under the law of the excluded middle

Here is a proof:

1. (-(-(x) & -(align)))
2. (~(-(x) & -(align)) + 1) by a
3. ((((-(x))) | ((-(align)))) + 1) by b
4. ((((((x) - 1))) | ((((align) - 1)))) + 1) by a
5. (((((x) - 1)) | (((align) - 1))) + 1) by c

Then if we just clean up the unnecessarily parenthesis, we arrive at @perfinion's code.

@behlendorf This looks good to me, provided the commit message includes the proof.

[perfinion]

I updated the commit message after discussion with @ryao and included the proof from the issue. Also added Reviewed by @ryao and @chrisrd.