Path: csiph.com!x330-a1.tempe.blueboxinc.net!usenet.pasdenom.info!aioe.org!feeder.news-service.com!kanaga.switch.ch!news-zh.switch.ch!switch.ch!news.belwue.de!not-for-mail
From: Thomas Richter <thor@math.tu-berlin.de>
Newsgroups: comp.lang.java.programmer
Subject: Re: Java left shift and right shift operators.
Date: Sat, 30 Apr 2011 00:02:15 +0200
Organization: InterNetNews at News.BelWue.DE (Stuttgart, Germany)
Lines: 44
Message-ID: <ipfch9$fto$1@news.belwue.de>
References: <295e16b3-2ed8-4529-bfb0-1cc26ed93ad6@d26g2000prn.googlegroups.com> <188cz6ta97kkm$.dlg@kimmeringer.de> <ip6igu$j0g$2@news.albasani.net> <aecb4896-0502-4d79-bc7d-c6ef2219ed94@r35g2000prj.googlegroups.com> <ip8a8m$iul$1@dont-email.me>
NNTP-Posting-Host: vpn-m-8d3a2f26.campus.uni-stuttgart.de
Mime-Version: 1.0
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Content-Transfer-Encoding: 7bit
X-Trace: news.belwue.de 1304114537 16312 141.58.47.38 (29 Apr 2011 22:02:17 GMT)
X-Complaints-To: news@news.belwue.de
NNTP-Posting-Date: Fri, 29 Apr 2011 22:02:17 +0000 (UTC)
User-Agent: Mozilla/5.0 (X11; U; Linux x86_64; en-US; rv:1.9.1.16) Gecko/20110307 Icedove/3.0.11
In-Reply-To: <ip8a8m$iul$1@dont-email.me>
Xref: x330-a1.tempe.blueboxinc.net comp.lang.java.programmer:3365

On 27.04.2011 07:40, Travers Naran wrote:
> On 26/04/2011 9:32 AM, Sanny wrote:
>> I have computed and found "if statments" take 20-50 times longer than
>> basic arithmetic operations. So I wanted to replace if condition by
>> some arithmetics/ Maths.
>
> This is why I think assembly language should be mandatory in schools. :-)
>
> An if statement gets translated into something like (using a
> pseudo-assembly for readability):
>
> ;; if (b > 0) return a << b else return a >> -b
> compare b to 0
> jump-less-than to L1
> a := a << b
> jump L2
> L1:
> b := -b
> a := a >> b
> L2:
> return ;; result in a

Folks, this is all too complicated. As already stated, branches can kill 
the performance, but it is just too easy without them:

(x >> r) << l.

And set l = max(-s,0) and r = max(s,0) where is is the "signed right 
shift". If the shift is always by the same constant, it is easier to 
precalculate l and r once and apply the shift all over again.

Alternatively, if the shift is always dynamic, l and r can also be 
computed without any branch:

int m = (s >> 31);
int r = s & m;
int l = (-s) & (~m);
int x = (x >> r) << l;

Whether this is faster depends of course whether a branch prediction is 
able to predict the branches well.

Greetings,
	Thomas