From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-return-93498-listarch-gcc=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 21667 invoked by alias); 26 Mar 2004 19:54:51 -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 20742 invoked from network); 26 Mar 2004 19:54:44 -0000
Received: from unknown (HELO mail0.lsil.com) (147.145.40.20)
  by sources.redhat.com with SMTP; 26 Mar 2004 19:54:44 -0000
Received: from milmhbs0.lsil.com (mhbs.lsil.com [147.145.1.30])
	by mail0.lsil.com (8.12.8/8.12.8) with ESMTP id i2QJsflW008238
	for <gcc@gcc.gnu.org>; Fri, 26 Mar 2004 11:54:41 -0800 (PST)
Received: from dts0.lsil.com (dts0.lsil.com [147.145.146.6])
	by milmhbs0.lsil.com (8.12.11/8.12.10) with ESMTP id i2QJuanL008538
	for <gcc@gcc.gnu.org>; Fri, 26 Mar 2004 11:56:36 -0800
Received: from dt219 (dt219 [147.145.146.219])
	by dts0.lsil.com (8.11.6+Sun/8.11.6) with SMTP id i2QJsfT27909
	for <gcc@gcc.gnu.org>; Fri, 26 Mar 2004 13:54:42 -0600 (CST)
Message-Id: <200403261954.i2QJsfT27909@dts0.lsil.com>
Date: Fri, 26 Mar 2004 22:21:00 -0000
From: John Lu <jlu@lsil.com>
Reply-To: John Lu <jlu@lsil.com>
Subject: Register Allocation
To: gcc@gcc.gnu.org
MIME-Version: 1.0
Content-Type: TEXT/plain; charset=us-ascii
Content-MD5: gKnPOdIRaZop6R8++oc4pg==
X-Scanned-By: MIMEDefang 2.39
X-SW-Source: 2004-03/txt/msg01579.txt.bz2

Hi,

I've been working on a port based on gcc-3.3 and I've noticed that
better assembly code is generated if the C source has
separate variables declared for distinct live ranges.

For example,

     int foo1(int *p1, int *p2) {
       int i;  /* one index variable for two live ranges */
       int total;

       total=0;
       for (i=0; i<100; i++) {
         total+=p1[i];
       }

       for (i=0; i<100; i+=2) {
         total+=p1[i];
       }
  
       return(total);
     }
     
produces worse code than:

     int foo2(int *p1, int *p2) {
       int i1,i2;  /* two index variables for two live ranges */
       int total;
   
       total=0;
       for (i1=0; i1<100; i1++) {
         total+=p1[i1];
       }

       for (i2=0; i2<100; i2+=2) {
         total+=p1[i2];
       }
  
       return(total);
     }
     
In my port, i1 is allocated a loop register which supports faster looping,
and i2 is allocated to a gpr because a decrement by two is needed.  When
only one index variable is declared, "i" is allocated to a loop register,
but this creates bad code for the second loop, since decrement by two is
not supported  for loop registers.  This happens with and without the
"-fnew-ra" option.

I was wondering if anyone else has seen this or am I doing something wrong
in my port.

Thanks,
John Lu