Re: [PATCH] Make sure the value in abs() does not get truncated if it is greater than 2^32

[Rolf Eike Beer]

Andrew Morton wrote:
> On Thu, 3 Sep 2009 23:12:01 +0200
> 
> Rolf Eike Beer <eike-kernel@xxxxxxxxx> wrote:
> > I was just digging a bit around in linux/kernel.h and stumbled over the
> > abs() makro. For me it looks as it would return wrong results on 64 bit
> > platforms if the input value is greater than 2^32.
> 
> Yes, it will truncate.
> 
> > diff --git a/include/linux/kernel.h b/include/linux/kernel.h
> > index d6320a3..1e6eb66 100644
> > --- a/include/linux/kernel.h
> > +++ b/include/linux/kernel.h
> > @@ -145,7 +145,7 @@ extern int _cond_resched(void);
> >  #define might_sleep_if(cond) do { if (cond) might_sleep(); } while (0)
> >
> >  #define abs(x) ({				\
> > -		int __x = (x);			\
> > +		long __x = (x);			\
> >  		(__x < 0) ? -__x : __x;		\
> >  	})
> 
> I wonder if that ends up producing worse code in the normal case.
> 
> We could use typeof to address that but what about
> 
> 	unsigned foo = -1;
> 	signed bar = abs(foo);
> 
> ?
> 
> That'll currently return 1 but if we use typeof it'll return
> 0xffffffffU which gets turned into -1.

I have not tested this, but what might happen with:

unsigned long a = 7;
long b = 9;
return abs(a - b);

I thought about using typeof but that's too much compiler magic for me to get 
right. If you can, go for it.

We could also try to check sizeof(__x) and then use either int or long to get 
the same code for the common code but behave right for big values.

Eike

Attachment: signature.asc
Description: This is a digitally signed message part.