Mailing-List:	contact cygwin-help AT cygwin DOT com; run by ezmlm
List-Subscribe:	<mailto:cygwin-subscribe AT cygwin DOT com>
List-Archive:	<http://sourceware.org/ml/cygwin/>
List-Post:	<mailto:cygwin AT cygwin DOT com>
List-Help:	<mailto:cygwin-help AT cygwin DOT com>, <http://sourceware.org/ml/#faqs>
Sender:	cygwin-owner AT cygwin DOT com
Mail-Followup-To:	cygwin AT cygwin DOT com
Delivered-To:	mailing list cygwin AT cygwin DOT com
Message-Id:	<6.0.1.1.0.20040718063702.01f796e8@imap.myrealbox.com>
X-Sender:	tprince AT imap DOT myrealbox DOT com
Date:	Sun, 18 Jul 2004 07:08:40 -0700
To:	cygwin AT cygwin DOT com
From:	Tim Prince <tprince AT computer DOT org>
Subject:	Re: Performance: g++ Cygwin vs. other compilers (copying char[] to vector)
In-Reply-To:	<cddooq$ssi$1@sea.gmane.org>
References:	<cddooq$ssi$1 AT sea DOT gmane DOT org>
Mime-Version:	1.0
X-IsSubscribed:	yes

At 05:00 AM 7/18/2004, Alex Vinokur wrote:

>Hi,
>
>How to explain so considerable difference in performance: g++ Cygwin vs. 
>other compilers in tests below?

I can't figure out at a glance what you are doing.  I find that loop 
lengths of 1000 or less (32-bit data types) require the cpu cycle timer 
(rdtsc) to get repeatable results.  I have no significant difference in 
timings between g++ STL templates running in linux or in cygwin, not even 
between 32-bit cygwin under WOW64 and 64-bit linux.  The Microsoft builds 
are generally slower.  I haven't tried -mno-cygwin, I've used only standard 
gcc/g++ 3.4.x.  My tests call a dummy() function before running back 
through the same test data, so that there is no software optimization 
performed by recognizing do-nothing loops.  The data are initialized to 
normal stuff of the correct data type, so that no exceptions occur, same 
initial values for each compiler.

Much of this comes ready-made in the Levine-Callahan-Dongarra vector 
benchmark from netlib.org.  Where possible, I have made C, C++, and Fortran 
90 versions, so I can get quite picky about effectiveness of compiler 
optimizations.

I don't find your compile options, or whether you have profiled.  For g++ 
under cygwin, I use
-O3 -funroll-loops -march=pentium4 -mfpmath=sse
No doubt, some of the compilers have entirely different defaults.  Are you 
one of those who considers the only fair comparison to be among defaults, 
even though those vary from no optimization to more than what is permitted 
by the language standards?

Microsoft STL doesn't implement copy() as memmove(), as libstdc++, STLport, 
and Dinkumware do, nor do the Microsoft compilers perform any consolidation 
of moves into 64- or 128- bit moves.  If there is a memmove() replacement, 
you should tell us which one is used.  Newlib is likely to be slow; even 
glibc or MS RT are not likely to be optimized for a particular style of CPU.



Tim Prince 


--
Unsubscribe info:      http://cygwin.com/ml/#unsubscribe-simple
Problem reports:       http://cygwin.com/problems.html
Documentation:         http://cygwin.com/docs.html
FAQ:                   http://cygwin.com/faq/

- Raw text -