Re: [PATCH] asm-generic/bitops/fls64.h

["Alexander van Heukelum" <heukelum@fastmail.fm>]

On Tue, 13 May 2008 14:58:39 +0100, "Russell King"
<rmk+lkml@arm.linux.org.uk> said:
> On Tue, May 13, 2008 at 03:24:13PM +0200, Alexander van Heukelum wrote:
> > On Tue, 13 May 2008 16:29:04 +0400, "Nickolay Vinogradov"
> > <nickolay@protei.ru> said:
> > > Alexander van Heukelum &#1087;&#1080;&#1096;&#1077;&#1090;:
> > > 
> > > > Hi Nickolay,
> > > > 
> > > > The change is ok, I guess, but the cast should be a no-op (fls
> > > > takes an int, which is always 32 bit in linux). What is the problem
> > > > you are seeing? Does fls64() return a wrong value in some cases? If
> > > > so, what cpu? Which values?
> > > > 
> > > > Why would this be a bug on big endian systems only? There is no
> > > > pointer magic involved, so the compiler should take care of the
> > > > casts in a correct way.
> > > > 
> > > > Maybe you see a compiler warning? Which compiler version?
> > > > 
> > > > (also note that current (development) kernels now have separate
> > > > versions for 32-bit and 64-bit environments.)
> > > 
> > > Because fls() is a macro for asm-arm:
> > > 
> > > #define fls(x) \
> > >          ( __builtin_constant_p(x) ? constant_fls(x) : \
> > >          ({ int __r; asm("clz\t%0, %1" : "=r"(__r) : "r"(x) : "cc"); 
> > > 32-__r; }) )
> > > 
> > > We can fix it right here:
> 
> No.  "fls" is for finding the last set bit in an _int_.  It is not
> supposed to have random crap passed to it, such as types longer than
> sizeof(int).
> 
> If you're going to pass long long (64-bit) arguments to fls, and then
> cast them to a u32, you're truncating the value, and you'll get the
> wrong answer if bit 33 or greater is set.  If you don't actually care
> about the upper bits, don't pass a 64-bit quantity to fls().
> 
> If you want to use fls with a long long, use fls64 instead.  Or for top
> marks, use a u64 and fls64.

But that was the problem we began with: the generic fls64 passes an u64
to fls. Nickolay's original patch solves that by putting a cast to u32
in fls64. I did not, however, understand why the cast was needed. It
should not be needed for correctness, imho, because fls is expected to
behave as if it was a function "int fls(int)", like ffs. Values passed
to fls should thus be truncated if necessary.

Making the truncation explicit in fls64 is still a good idea, though.

Greetings,
    Alexander

> -- 
> Russell King
>  Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
>  maintainer of:
-- 
  Alexander van Heukelum
  heukelum@fastmail.fm

-- 
http://www.fastmail.fm - The way an email service should be