Subj : Re: Multiplications faster than bit shifts?
To   : comp.programming
From : websnarf
Date : Sat Jul 09 2005 04:08 pm

Willem wrote:
> Eric wrote:
> ) I read a post by someone recently, and they asserted that multiplications
> ) by two were faster on recent x86 cpus than simple bit shifts.
> )
> ) I have a very hard time believing a simple shift would take longer than a
> ) multiplication and wanted to know what you people thought about this.
>
> A multiplication by two is (probably) slower than a shift by one bit, but
> a multiplication by sixteen is probably faster than a shift by four bits.
>
> This is because recent x86 cpus don't have barrel shifters anymore,
> so they have to do X consecutive shifts to get a shift by X bits.

You better define what you mean by "recent x86 CPUs".  The Athlon and
Athlon 64s have 3 parallel barrel shifters.  And Intel has all but
abandoned the Pentium 4 (where they had no barrel shifter) and gone
back to the Pentium-Pro architecture (now dubbed Pentium-M) which has
one barrel shifter.

> I'm not sure where the boundary is between a shift or a multiply
> being faster.

For powers of two, even on the Pentium 4, shifting is faster than
multiplication.

BTW, shift by one is optimally replaced by adding the value to itself.

-- 
Paul Hsieh
http://www.pobox.com/~qed/
http://bstring.sf.net/

.