Re: memcpy is not optimal implemented

[Andrew Morton]

Ulrich Drepper wrote:
> 
> "Wallner, Jens" <address@hidden> writes:
> 
> > the currend memcpy function for i686 is not optimal implemented.
> 
> That's not correct.  The underlying assumptions are different. 

On Intel chips, `rep;movsd' is *slower* than an open-coded
copy loop for all alignments except when source and dest
are both 8-byte-aligned.

When src and dest are 8-byte-aligned, movsd is about 20%
faster.  When they're not, an open-coded copy can be
up to 50% faster, depending upon alignment.

On a K6-2, movsd is faster at all alignments.

These are _huge_ differences, and I suggest they outweigh
this sort of algorithmic tweaking.

There's a little app for playing with this at

        http://www.uow.edu.au/~andrewm/linux/cptimer.tar.gz

if anyone's interested.