Re: More diff performance notes

Greg Hudson wrote:

> I wrote:
>
>>> (While I'm here, I have determined that I just about match
>>> diff+gzip on the .elc data set if I change the svndiff encoding so
>>> that new data immediately follows "copy from new data" instructions
>>> instead of being stored in a separate part of the window. The
>>> average encoded bytes per instruction drops to 2.48 in the binary
>>> case and 2.80 in the text case.)
>>
>
> Branko wrote:
>
>> Hmmm? Why would that help? The size of the new data doesn't change
>> by moving around, does it? Where's the savings in that, then?
>
>
> You no longer need an offset when you encode "copy from new data"
> instructions.

But you never need it, anyway. The offset in the new
data is implicit -- you start at 0 and advance it by
<size> in every add instruction.

-- 
Brane ?Cibej
home:   <brane_at_xbc.nu> http://www.xbc.nu/brane/
ACM:   <brane_at_acm.org> http://www.acm.org/