From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-return-65416-listarch-gcc=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 29924 invoked by alias); 30 Dec 2002 18:20:42 -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 29894 invoked from network); 30 Dec 2002 18:20:41 -0000
Received: from unknown (HELO tone.orchestra.cse.unsw.EDU.AU) (129.94.242.28)
  by 209.249.29.67 with SMTP; 30 Dec 2002 18:20:41 -0000
Received: From cse.unsw.edu.au ([129.94.241.37] == ppp2-037.ppp2.cse.unsw.EDU.AU)
	(for <gcc@gcc.gnu.org>) By tone With Smtp ;
	Tue, 31 Dec 2002 05:20:28 +1100 
From: Qiong Cai <qiongc@cse.unsw.edu.au>
To:  gcc@gcc.gnu.org
Date: Mon, 30 Dec 2002 12:26:00 -0000
Message-ID: <3E112934.8060803@cse.unsw.edu.au>
Reply-To:  qiongc@cse.unsw.edu.au
Organization: CSE, UNSW
User-Agent: Mozilla/5.0 (X11; U; Linux i686; en-US; rv:1.0.0) Gecko/20020529
X-Accept-Language: en-us, en
MIME-Version: 1.0
Subject: gcse fails in this simple case
Content-Type: text/plain; charset=us-ascii; format=flowed
Content-Transfer-Encoding: 7bit
X-SW-Source: 2002-12/txt/msg01590.txt.bz2

Hi,

I have the follwing code to test the implemenation of
GCSE optimization in GCC.

0: int a, b;
1:
2: int fun(int ii, int jj)
3: {
4:   int sum1, sum2;
5:   sum1 = sum2 = 0;
6:
7:   if ( ii > jj )
8:     sum1 = a + b;
9:   else
10:   sum2 = a + b;
11:
12:  return sum1+sum2;
13: }

However, GCC cannot find the common subexpression
at Line 8 and 10. From the *.gcse debug dump,
I found GCC consider the expression a+b as two different
expressions at RTL level. The following RTL codes
are from *.gcse dump:


insn 22 57 24 (set (reg:SI 63)
         (mem/f:SI (symbol_ref:SI ("a")) [2 a+0 S4 A32])) 45 {*movsi_1}
...

(insn 24 22 26 (set (reg:SI 64)
         (mem/f:SI (symbol_ref:SI ("b")) [2 b+0 S4 A32])) 45 {*movsi_1} ...

(insn 26 24 27 (parallel[               <= a+b at Line 8
             (set (reg/v:SI 61)
                 (plus:SI (reg:SI 63)
                     (reg:SI 64)))
             (clobber (reg:CC 17 flags))
         ] ) 207 {*addsi_1} (nil)
     (nil))


...
(insn 32 58 34 (set (reg:SI 65)
         (mem/f:SI (symbol_ref:SI ("a")) [2 a+0 S4 A32])) 45 {*movsi_1}
...
(insn 34 32 36 (set (reg:SI 66)
         (mem/f:SI (symbol_ref:SI ("b")) [2 b+0 S4 A32])) 45 {*movsi_1}
....

(insn 36 34 37 (parallel[            <= a + b at Line 10
             (set (reg/v:SI 62)
                 (plus:SI (reg:SI 65)
                     (reg:SI 66)))
             (clobber (reg:CC 17 flags))
         ] ) 207 {*addsi_1} (nil)
     (nil))

 From the above codes, we see gcse fails to eliminate
the common loads of "a" and "b", and then fails to eliminate "a+b".


I think it is a bug, or I miss something.

My GCC compiler:
 > gcc -v
 > Reading specs from /usr/local/lib/gcc-lib/i686-pc-linux-gnu/3.2/specs
 > Configured with: ../gcc/configure --disable-libgcj
 > Thread model: posix
 > gcc version 3.2


Qiong