[PATCH v3] linux/kernel.h: Fix DIV_ROUND_CLOSEST to support negative operands

[PATCH v3] linux/kernel.h: Fix DIV_ROUND_CLOSEST to support negative operands

[Guenter Roeck]

DIV_ROUND_CLOSEST returns a bad result for dividends with different sign:
	DIV_ROUND_CLOSEST(-2, 2) = 0

Most of the time this does not matter. However, in the hardware monitoring
subsystem, DIV_ROUND_CLOSEST is sometimes used on integers which can be
negative (such as temperatures).

Signed-off-by: Guenter Roeck <linux@roeck-us.net>
---
v3: Instead of adding a new macro, fix DIV_ROUND_CLOSEST.
    This version works for negative dividend and divisor.

v2: v1 did not work if typeof(divisor) was an unsigned variable type
    (which can obviously not be negative).
    Rework to revert to DIV_ROUND_CLOSEST if the dividend is unsigned,
    or if it is signed but non-negative.

 include/linux/kernel.h |    7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/include/linux/kernel.h b/include/linux/kernel.h
index 6043821..4b180de 100644
--- a/include/linux/kernel.h
+++ b/include/linux/kernel.h
@@ -84,8 +84,11 @@
 )
 #define DIV_ROUND_CLOSEST(x, divisor)(			\
 {							\
-	typeof(divisor) __divisor = divisor;		\
-	(((x) + ((__divisor) / 2)) / (__divisor));	\
+	typeof(x) __x = x;				\
+	typeof(divisor) __d = divisor;			\
+	((__x) < 0) == ((__d) < 0) ?			\
+		(((__x) + ((__d) / 2)) / (__d)) :	\
+		(((__x) - ((__d) / 2)) / (__d));	\
 }							\
 )
 
-- 
1.7.9.7

Re: [PATCH v3] linux/kernel.h: Fix DIV_ROUND_CLOSEST to support negative operands

[Andrew Morton]

On Thu, 30 Aug 2012 17:10:47 -0700 Guenter Roeck <linux@roeck-us.net> wrote:

> DIV_ROUND_CLOSEST returns a bad result for dividends with different sign:
> 	DIV_ROUND_CLOSEST(-2, 2) = 0
> 
> Most of the time this does not matter. However, in the hardware monitoring
> subsystem, DIV_ROUND_CLOSEST is sometimes used on integers which can be
> negative (such as temperatures).
> 
> ...
>
> --- a/include/linux/kernel.h
> +++ b/include/linux/kernel.h
> @@ -84,8 +84,11 @@
>  )
>  #define DIV_ROUND_CLOSEST(x, divisor)(			\
>  {							\
> -	typeof(divisor) __divisor = divisor;		\
> -	(((x) + ((__divisor) / 2)) / (__divisor));	\
> +	typeof(x) __x = x;				\
> +	typeof(divisor) __d = divisor;			\
> +	((__x) < 0) == ((__d) < 0) ?			\
> +		(((__x) + ((__d) / 2)) / (__d)) :	\
> +		(((__x) - ((__d) / 2)) / (__d));	\
>  }							\
>  )

Your v2 had that sneaky little "(typeof(x))-1 >= 0" trick in it, so
half the code gets elided at compile time if `x' (why isn't this called
"dividend") has an unsigned type.

Would retaining that be of any benefit?  We do want to avoid doing the
compare-and-branch in as many cases as possible.

Also, this would be a great opportunity to document the macro's beahviour
(I do go on).  That would be a useful thing to do, given that we're now
handling the four +/+, +/-, -/+, -/- cases and the behaviour for each
case isn't terribly obvious.

Re: [PATCH v3] linux/kernel.h: Fix DIV_ROUND_CLOSEST to support negative operands

[Guenter Roeck]

On Thu, Aug 30, 2012 at 05:35:31PM -0700, Andrew Morton wrote:
> On Thu, 30 Aug 2012 17:10:47 -0700 Guenter Roeck <linux@roeck-us.net> wrote:
> 
> > DIV_ROUND_CLOSEST returns a bad result for dividends with different sign:
> > 	DIV_ROUND_CLOSEST(-2, 2) = 0
> > 
> > Most of the time this does not matter. However, in the hardware monitoring
> > subsystem, DIV_ROUND_CLOSEST is sometimes used on integers which can be
> > negative (such as temperatures).
> > 
> > ...
> >
> > --- a/include/linux/kernel.h
> > +++ b/include/linux/kernel.h
> > @@ -84,8 +84,11 @@
> >  )
> >  #define DIV_ROUND_CLOSEST(x, divisor)(			\
> >  {							\
> > -	typeof(divisor) __divisor = divisor;		\
> > -	(((x) + ((__divisor) / 2)) / (__divisor));	\
> > +	typeof(x) __x = x;				\
> > +	typeof(divisor) __d = divisor;			\
> > +	((__x) < 0) == ((__d) < 0) ?			\
> > +		(((__x) + ((__d) / 2)) / (__d)) :	\
> > +		(((__x) - ((__d) / 2)) / (__d));	\
> >  }							\
> >  )
> 
> Your v2 had that sneaky little "(typeof(x))-1 >= 0" trick in it, so
> half the code gets elided at compile time if `x' (why isn't this called
> "dividend") has an unsigned type.
> 
> Would retaining that be of any benefit?  We do want to avoid doing the
> compare-and-branch in as many cases as possible.
> 
DIV_ROUND_CLOSEST(0,-2)=1

This also happens if I keep the sneaky code. The v3 code does not have this
problem. I know it is a bit theoretic, but still there. Of course, I could
simply ignore the divisor's sign entirely, assuming (and documenting) that
negative divisors are just too odd to deal with. Commentss welcome ...

> Also, this would be a great opportunity to document the macro's beahviour
> (I do go on).  That would be a useful thing to do, given that we're now
> handling the four +/+, +/-, -/+, -/- cases and the behaviour for each
> case isn't terribly obvious.
> 
Ok.

Guenter