From: grenie AT math DOT jussieu DOT fr
Date: Sun, 3 Oct 1999 16:34:28 +0200 (MET DST)
Message-Id: <199910031434.QAA10824@riemann.math.jussieu.fr>
To: pgcc AT delorie DOT com
X-URL: http://goof.com/pcg/questions.html
X-Mailer: Lynx, Version 2.7
X-Personal_name: Loïc Grenié
Subject: Optimization pug in pgcc-2.95.1
Cc: grenie AT math DOT jussieu DOT fr
MIME-Version: 1.0
Content-Type: text/plain; charset=iso-8859-1
Content-Transfer-Encoding: 8bit
Reply-To: pgcc AT delorie DOT com

     Hello,

     I've found a bug in pgcc-2.95.1. When I compile the following program:

func(void)
{
    int tab[512], i;

    for (i = 0; i < 511; i++)
        tab[i+1] = tab[i];
}

  with gcc -O4 -mpentium I obtain code which is incorrect. Namely, at the end
  of the loop the pointer &tab[i] and the variable i are incremented 512
  times two much. Here is the code I obtain:

func:
        pushl %ebp
        movl %esp,%ebp
        subl $2064,%esp
        movl $-511,%ecx
        pushl %esi
        pushl %ebx
        leal -2048(%ebp),%edx
        movl $4,%esi
        xorl %ebx,%ebx
        .align 4
.L6:
        movl (%ebx,%edx),%eax
        movl %eax,(%esi,%edx)
        addl $4,%esi
        addl $2048,%ebx
        addl $512,%ecx
        jne .L6
        popl %ebx
        popl %esi
        movl %ebp,%esp
        popl %ebp
        ret

  the addl $2048,%ebx should be a addl $4,%ebx and the addl $512,%ecx should
  be addl $1,%ecx. I've tried to isolate the bug in the sources, but I've
  failed...

    If the correction is small, I'd like to have it as soon as it's been made
  by email if possible.

        Loïc

PS: another idea for optimization: the following code would be better:

func:
        pushl %ebp
        movl %esp,%ebp
        subl $2064,%esp
        movl $-511,%ecx
	pushl %esi
        pushl %ebx
        leal -2048(%ebp),%edx
        xorl %ebx,%ebx
        .align 4
.L6:
        movl (%ebx,%edx),%eax
        addl $4,%ebx
        movl %eax,(%esi,%edx)
        addl $4,%esi
        addl $1,%ecx
        jne .L6
        popl %ebx
	popl %esi
        movl %ebp,%esp
        popl %ebp
        ret

  and %esi could be eliminated by using only %ebx, but I'm not sure that would
  be really interesting in this setup.