Re: [PATCH] iio: inkern: Avoid risky abs() usage in iio_multiply_value()

[Andy Shevchenko <andriy.shevchenko@intel.com>]

On Tue, Mar 31, 2026 at 04:26:35PM +0100, David Laight wrote:
> On Tue, 31 Mar 2026 12:29:22 +0300
> Andy Shevchenko <andriy.shevchenko@intel.com> wrote:
> > On Tue, Mar 31, 2026 at 10:49:59AM +0200, Romain Gantois wrote:

> > > iio_multiply_value() passes integers val and val2 directly to abs(). This
> > > is problematic because if a signed argument to abs is the lowest value for
> > > its type, then the result is undefined due to overflow.
> > > 
> > > Cast val and val2 to s64 before passing them to abs() to avoid this issue.  

...

> I've just looked at the 'work of art' that is abs().
> What is wrong with:
> #define abs(x) (sizeof(x) == sizeof(long long) ? __abs(long long, x) : \
> 		__abs(int, x))
> #define __abs(type, x) \
> 	({ type __abs_x = (x); __abs_x < 0 ? -__abs_x : __abs_x;})
> 
> It is just as broken for u128.
> It will use the correct signedness for char (but it is unsigned now).
> It doesn't cast back to char, but that is entirely pointless unless code
> looks at the type of the expression, the return value itself is always
> promoted to int before being used.
> 
> Actually replace the -__abs_x (UB for INT_MIN) with the safe:
> 	(unsigned type)-(__abs_x + 1) + 1
> and the return type will be unsigned with a correct value for -INT_MIN.
> (Oh and the compiler sees through the mess.)

And this is definitely wrong. We must keep type, because abs() might be used in
the comparisons with signed or as parameter to multiplication or division where
sign has to be preserved.


-- 
With Best Regards,
Andy Shevchenko