Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Tue, 26 Jun 2012 09:24:15 +0200
From: magnum <>
Subject: Re: md5cryptsse()'s uses of length[]

On 2012-06-26 05:41, Solar Designer wrote:
> On Tue, Jun 26, 2012 at 06:13:13AM +0400, Solar Designer wrote:
>> In sse-intrinsics.c: md5cryptsse(), there's a loop that assigns password
>> lengths to length[i], then uses length[i] many times.  Wouldn't it be
>> better to use an int variable local to that loop instead, and only at
>> the end of each loop iteration store the value into length[i]?
> Attached is a patch that fixes this.  No measurable performance
> difference from that (since it's out of the 1000 iterations loop).
> However, while at it I also spotted the integer division in dispatch()
> (I am surprised: even the old non-vectorized code avoided that), which I
> patched out for a 0.5% speedup here.
> We might want to regenerate and re-benchmark icc's .S files now.

Ouch, and my icc is not working right now. Simon, could you pull 
magnum-jumbo, do a "make intrinsics" and post the resulting files?


Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.