Re: bridge: faster compare for link local addresses

[Eric Dumazet <dada1@cosmosbay.com>]

On Tuesday 13 March 2007 15:01, Andi Kleen wrote:
> David Miller <davem@davemloft.net> writes:
> > From: Rick Jones <rick.jones2@hp.com>
> > Date: Mon, 12 Mar 2007 17:05:39 -0700
> >
> > > Being paranoid - are there no worries about the alignment of dest?
> >
> > If it's an issue, it's an issue elsewhere too, as the places
> > where Stephen took this idiomatic code from is the code
> > ethernet handling and that runs on every input packet via
> > eth_type_trans().
>
> As a quick note -- when you tell gcc the expected alignment
> by using correct types then moderm gcc should generate fast inline code
> for memcpy/memcmp/etc. by itself. It only falls back to a slow generic
> function when it cannot figure out the alignment or the size.
>
> So I expect just using u32 * instead of char * should have the same
> effect and would be somewhat cleaner and the memcmp could be kept.

For memcpy() yes you can have some optimizations.

But memcmp() has a strong semantic (in libc). memcmp(a, b, 6) should do 6 byte 
compares and conditional branches, regardless of a/b alignment.
Or use the x86 "rep cmpsb" instruction that basically has the same cost.

The trick we use in compare_ether_addr() reduces to one some arithmetic and 
one test.

return ((a[0] ^ b[0]) | (a[1] ^ b[1]) | (a[2] ^ b[2])) != 0;

I found this line as clean as memcmp(a, b, 6)

(On x86_64, were alignment is not mandatory, we could do :

((*(long *)a ^ *(long*)b) << 16) != 0)

(only if we can always read two extra bytes without faulting, of course :) )