[Bug rtl-optimization/56151] [4.8 Regression] Performance degradation after r194054 on x86 Atom.

[ebotcazou at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56151

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |ebotcazou at gcc dot
                   |                            |gnu.org

--- Comment #5 from Eric Botcazou <ebotcazou at gcc dot gnu.org> 2013-02-04 10:34:25 UTC ---
> Patch to help explain the problem:
> 
> Index: optabs.c
> ===================================================================
> --- optabs.c    (revision 195687)
> +++ optabs.c    (working copy)
> @@ -1452,8 +1452,13 @@ expand_binop_directly (enum machine_mode
>      {
>        /* If PAT is composed of more than one insn, try to add an appropriate
>          REG_EQUAL note to it.  If we can't because TEMP conflicts with an
> -        operand, call expand_binop again, this time without a target.  */
> +        operand, call expand_binop again, this time without a target.
> +
> +        However, if target is a MEM, just accept the lossage of not having
> +        a REG_EQUAL note.  This avoids splitting up insns of the form
> +        "MEM=MEM op X", a form we may not be able to reconstruct later.  */
>        if (INSN_P (pat) && NEXT_INSN (pat) != NULL_RTX
> +         && ! (target && MEM_P (target))
>           && ! add_equal_note (pat, ops[0].value, optab_to_code (binoptab),
>                                ops[1].value, ops[2].value))
>         {

I agree that we should try to do something like that.  Generating inferior code
to be able to add a REG_EQUAL that doesn't help in the end isn't very clever.
There is also the same pattern in expand_unop_direct.

Is that the "maximal" test that can be added to catch this case or could it be
made more precise?