Re: [PATCH] cciss: Fix warnings during compilation under 32bitenvironment

[Andrew Morton <akpm@linux-foundation.org>]

On Fri, 20 Apr 2007 16:20:59 -0400
James Bottomley <James.Bottomley@SteelEye.com> wrote:

> On Fri, 2007-04-20 at 12:30 -0700, Andrew Morton wrote:
> > On Fri, 20 Apr 2007 14:50:06 -0400
> > James Bottomley <James.Bottomley@SteelEye.com> wrote:
> > 
> > > > CONFIG_LBD=y gives us an additional 3kb of instructions on i386
> > > > allnoconfig.  Other architectures might do less well.  It's not a huge
> > > > difference, but that's the way in which creeping bloatiness happens.
> > > 
> > > OK, sure, but if we really care about this saving, then unconditionally
> > > casting to u64 is therefore wrong as well ... this is starting to open
> > > quite a large can of worms ...
> > > 
> > > For the record, if we have to do this, I fancy sector_upper_32() ... we
> > > should already have some similar accessor for dma_addr_t as well.
> > 
> > hm.  How about this?
> > 
> > --- a/include/linux/kernel.h~upper-32-bits
> > +++ a/include/linux/kernel.h
> > @@ -40,6 +40,17 @@ extern const char linux_proc_banner[];
> >  #define DIV_ROUND_UP(n,d) (((n) + (d) - 1) / (d))
> >  #define roundup(x, y) ((((x) + ((y) - 1)) / (y)) * (y))
> >  
> > +/**
> > + * upper_32_bits - return bits 32-63 of a number
> > + * @n: the number we're accessing
> > + *
> > + * A basic shift-right of a 64- or 32-bit quantity.  Use this to suppress
> > + * the "right shift count >= width of type" warning when that quantity is
> > + * 32-bits.
> > + */
> > +#define upper_32_bits(n) (((u64)(n)) >> 32)
> 
> Won't this have the unwanted side effect of promoting everything in a
> calculation to long long on 32 bit platforms, even if n was only 32
> bits?

bummer.

> > +
> > +
> >  #define	KERN_EMERG	"<0>"	/* system is unusable			*/
> >  #define	KERN_ALERT	"<1>"	/* action must be taken immediately	*/
> >  #define	KERN_CRIT	"<2>"	/* critical conditions			*/
> > _
> > 
> > It seems to generate the desired code.  I avoided Alan's ((n >> 31) >> 1)
> > trick because it'll generate peculiar results with signed 64-bit
> > quantities.
> 
> I've seen the trick done similarly with ((n >> 16) >> 16) which
> shouldn't have the issue.

That works if we know the caller is treating the return value as 32 bits,
but we don't know that.

If we have

#define upper_32_bits(x)  ((x >> 16) >> 16)

then

	upper_32_bits(0x8888777766665555)

will return 0x88887777 if it's treated as 32-bits, but it'll return
0xffffffff88887777 if the caller is using 64-bits.

I spose

#define upper_32_bits(x)  ((u32)((x >> 16) >> 16))

will do the trick.