From mboxrd@z Thu Jan  1 00:00:00 1970
Return-Path: <gcc-return-101839-listarch-gcc=gcc.gnu.org@gcc.gnu.org>
Received: (qmail 26807 invoked by alias); 1 Sep 2004 10:23:02 -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 26728 invoked from network); 1 Sep 2004 10:22:59 -0000
Received: from unknown (HELO smtp1.libero.it) (193.70.192.51)
  by sourceware.org with SMTP; 1 Sep 2004 10:22:59 -0000
Received: from localhost (172.16.1.79) by smtp1.libero.it (7.0.027-DD01)
        id 40C7315F0102095D; Wed, 1 Sep 2004 12:22:57 +0200
Received: from bagio (151.37.78.191) by smtp1.libero.it (7.0.027-DD01)
        id 40CB29090361FC64; Wed, 1 Sep 2004 12:22:56 +0200
Message-ID: <031701c4900d$a58ccfe0$bf4e2597@bagio>
From: "Giovanni Bajo" <giovannibajo@libero.it>
To: "Rakesh Kumar, Noida" <rakesh.kumar@noida.hcltech.com>
Cc: <gcc@gcc.gnu.org>
References: <33BC33A9E76474479B76AD0DE8A1697202C38D@exch-ntd.nec.noida.hcltech.com>
Subject: Re: stack height reduction in tree-SSA
Date: Wed, 01 Sep 2004 10:23:00 -0000
MIME-Version: 1.0
Content-Type: text/plain;
	charset="iso-8859-1"
Content-Transfer-Encoding: 7bit
X-Virus-Scanned: by amavisd-new at libero.it
X-SW-Source: 2004-09/txt/msg00026.txt.bz2

Rakesh Kumar, Noida wrote:

> I feel that this problem could have been fixed in tree-SSA, but
> exactly opposite is done. [...]
> Is it that good to forget the scope of variable and allocate each
> one a new stack slot? I'm unable to get any reasons why the stack
> slots can't be shared. Please make me aware of the issues involved.

Yes. tree-ssa behaves very bad wrt stack allocation. There is at least an open
regression filed by me about it (http://gcc.gnu.org/PR16987). The bug that
tracks this problem is http://gcc.gnu.org/PR9997. Notice that, from what I
heard, the correct solution would be to share stack slots using the live ranges
of the variables, rather than the mere lexical scopes. For instance:

void foo(void)
{
   int a, b;

   a = get();
   use_it(&a);

   b = get();
   use_it(&b);
}

could be optimized to use only one variable on the stack. I have the secret
hope that someone will eventually take care of this before 3.5 is release, but
it's just a hope. Surely, this bug alone regresses GCC badly for all embedded
systems.

Giovanni Bajo