From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-return-105986-listarch-gcc=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 1255 invoked by alias); 29 Nov 2004 23:54:30 -0000
Mailing-List: contact gcc-help@gcc.gnu.org; run by ezmlm
Precedence: bulk
List-Archive: <http://gcc.gnu.org/ml/gcc/>
List-Post: <mailto:gcc@gcc.gnu.org>
List-Help: <http://gcc.gnu.org/ml/>
Sender: gcc-owner@gcc.gnu.org
Received: (qmail 969 invoked from network); 29 Nov 2004 23:54:15 -0000
Received: from unknown (HELO avtrex.com) (67.116.42.149)
  by sourceware.org with SMTP; 29 Nov 2004 23:54:15 -0000
Received: from avtrex.com ([192.168.0.111] RDNS failed) by avtrex.com with Microsoft SMTPSVC(5.0.2195.6713);
	 Mon, 29 Nov 2004 15:54:14 -0800
Message-ID: <41ABB6B2.7020502@avtrex.com>
Date: Tue, 30 Nov 2004 00:24:00 -0000
From: David Daney <ddaney@avtrex.com>
User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.4.1) Gecko/20031030
MIME-Version: 1.0
To: Morten Welinder <terra@gnome.org>
CC:  gcc@gcc.gnu.org
Subject: Re: INT_MIN % -1
References: <20041129215914.A03FC1422D5B@darter.rentec.com>
In-Reply-To: <20041129215914.A03FC1422D5B@darter.rentec.com>
Content-Type: text/plain; charset=us-ascii
Content-Transfer-Encoding: 7bit
X-OriginalArrivalTime: 29 Nov 2004 23:54:14.0391 (UTC) FILETIME=[BAEAC070:01C4D66E]
X-SW-Source: 2004-11/txt/msg01179.txt.bz2

Morten Welinder wrote:
> What is the value of  INT_MIN % -1  supposed to be, assuming 32-bit
> ints and two-complement representation.
> 
> C99 seems to be telling us two things about '%':
> 
> (1) It computes remainder for integer division.
> (2) "If the quotient  a/b  is representable, the expression
>     (a/b)*b + a%b  shall equal  a."
> 
> (2) is not useful for the case since the quotient is not representable.  I'd
> say that (1) means that the result should be zero, but that would be rather
> expensive to implement as you would have to test for -1 at runtime on i386,
> for example.
> 
> I would claim that there is no overflow in the calculation: both arguments
> and the result are perfectly representable as 32-bit ints.  The fact that
> the quotient of the same values, if performed, overflows is no more relevant
> that the fact that you can't take the square root of the two sides.
> 
> gcc 3.4 on solaris/sparc seems to get zero; gcc 3.3.1 on linux gives me a
> crash at runtime.  (Because the signed integer division instruction traps
> as documented.)
> 
> Comments?
> 
> Morten
> 
> 
> 
> -----------------------------------------------------------------------------
> 
> 
>>uname -a
> 
> SunOS troll 5.8 Generic_117350-11 sun4u sparc
> 
>>./a.out 
> 
> -2147483648 % -1 (const) = 0
> -2147483648 % -1 (non-const) = 0
> 
> -----------------------------------------------------------------------------
> 
> 
>>uname -a
> 
> Linux darter 2.4.21-243-default #1 Thu Aug 12 15:22:14 UTC 2004 i686 i686 i386 GNU/Linux
> 
>>./a.out 
> 
> -2147483648 % -1 (const) = 0
> Floating point exception
> 
> -----------------------------------------------------------------------------

/junk # uname -a
Linux (none) 2.4.25-Avtrex #152 Mon Nov 15 10:51:36 PST 2004 mips unknown
/junk # ./bla
-2147483648 % -1 (const) = 0
-2147483648 % -1 (non-const) = 0

Intel chose to trap.  Sun and MIPS chose not to.

I don't have the spec. in front of me, but I can imagine that it allows for
this.

If you want well defined behavior use something like java.  The JLS
specifies exactly what should happen in this case.

David Daney.