Re: Loop optimizer misses simple optimisation?

[Kamil Iskra <kamil@dwd.interkom.pl>]

On Wed, 29 Apr 1998, Jim Wilson wrote:

> 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.

I see. But please have a look at udivmodsi4 in m68k.md:

(define_insn "udivmodsi4"
  [(set (match_operand:SI 0 "general_operand" "=d")
	(udiv:SI (match_operand:SI 1 "general_operand" "0")
		 (match_operand:SI 2 "general_operand" "dmsK")))
   (set (match_operand:SI 3 "general_operand" "=d")
	(umod:SI (match_dup 1) (match_dup 2)))]
  "TARGET_68020 && !TARGET_5200"
  "*
{
  if (find_reg_note (insn, REG_UNUSED, operands[3]))
    return \"divu%.l %2,%0\";
  else
    return \"divul%.l %2,%3:%0\";
}")

I don't know enough about division instructions available in m68020+, but
it looks to me from the above that there is an instruction that doesn't
generate the reminder, and that it is actually generated when the reminder
is not needed. Wouldn't it be better to move it to a separate, new udivsi3
pattern, instead of relying on REG_UNUSED, which is only available when
optimising, I suppose? This would allow the LICM code to do a better job,
wouldn't it?

/ Kamil Iskra    AmigaOS  Linux/i386  Linux/m68k               \
| GeekGadgets GCC maintainer   UNIX system administrator       |
| iskra@student.uci.agh.edu.pl  kiskra@ernie.icslab.agh.edu.pl |
\ kamil@dwd.interkom.pl   http://student.uci.agh.edu.pl/~iskra /