From mboxrd@z Thu Jan  1 00:00:00 1970
From: Jim Wilson <wilson@cygnus.com>
To: Kamil Iskra <kiskra@ernie.icslab.agh.edu.pl>
Cc: egcs@cygnus.com
Subject: Re: Loop optimizer misses simple optimisation? 
Date: Wed, 29 Apr 1998 15:21:00 -0000
Message-id: <199804292128.OAA01108@rtl.cygnus.com>
References: <Pine.GSO.3.95.980428154521.25424H-100000@ernie>
X-SW-Source: 1998-04/msg01166.html

I checked, and Mike Meissner's comment is correct.  The problem here is that
the m68k divide instruction computes two values (quotient and remainder),
giving us:
(insn 83 33 84 (parallel[
            (set (reg:SI 34)
                (udiv:SI (const_int 4194304)
                    (reg/v:SI 29)))
            (set (reg:SI 35)
                (umod:SI (const_int 4194304)
                    (reg/v:SI 29)))
        ] ) -1 (nil)
    (nil))

In scan_loop in loop.c, the code that recognizes loop invariant instructions
does this:

      if (GET_CODE (p) == INSN
          && (set = single_set (p))
          && GET_CODE (SET_DEST (set)) == REG
          && ! may_not_optimize[REGNO (SET_DEST (set))])

The single_set checks means that insns that compute two values are never
considered for loop-invariant-code-motion, hence we won't move a m68k
divide out of a loop.

We would need to generalize the LICM code to handle insns with multiple
sets to fix this.

Jim