I'm resending the patch as is and adding what I hope are the right CCs.  Also 
let me explain why I opted to add abs64() and use the gcc builtin.

quoted text
>Can't we just improve abs? Say,

I was reluctant to change abs() since it would have a much larger impact on 
the code base.  Using typeof() should be OK but if any of the callers 
mistakenly call abs() with an unsigned value then we could see compiler 
warnings about '__x < 0' being a useless conditional.

quoted text
>This is a bit unusual. I mean, it is not that common to use gcc builtins
>in the normal code. And, it seems, we can use __fls(divisor >> 32) or
>just fls64() instead ?

I opted for the gcc builtin because I felt it made the code more readable.  I 
also suspect it will perform slightly better than __fls() on some archs.  For 
example, on powerpc __fls() in implemented using the 'cntlzw' instruction.  
It returns (BITS_PER_LONG - 1 - cntlzw) which is wasted work since my 
function would immediately undo this to get back cntlzw.  If I was lucky the 
compiler would optimize this away for me but if I use the builtin I don't 
need to take the chance.

-- 
Thanks,
Brian Behlendorf