Re: [dataflow]: PATCH committed to fix another libcall issue in dce.

[Dorit Nuzman <DORIT at il dot ibm dot com>]

> This time for sure!!!!
>
> This patch allows dce to properly delete noop moves from the inside of
> libcalls.
> This appears to be the last regression for the dataflow branch on the
> sony ps3, but pinskia cannot do a full bootstrap and regtest until he
> returns from holiday.
>

I successfully built the branch on the SPU, and checked that 930513-2.c,
which ICE'd for Andrew (http://gcc.gnu.org/ml/gcc/2006-12/msg00279.html),
doesn't ICE. Also sent a full make check run - will let you know when it
ends.

dorit

> This does not add any other regressions to x86-64 or ppc.  Until the
> next merge, we cannot test on ia-64.
>
>
> 2006-12-25  Kenneth Zadeck <zadeck@naturalbridge.com>
>     * dce.c (delete_corresponding_reg_eq_notes): Added comment.
>     (delete_unmarked_insns): Added code to delete noop moves
>     inside of libcalls.  Changed to used delete_insn_and_edges.
>
> Index: dce.c
> ===================================================================
> --- dce.c   (revision 120165)
> +++ dce.c   (working copy)
> @@ -204,6 +204,10 @@ delete_corresponding_reg_eq_notes (rtx i
>       rtx note = find_reg_note (noted_insn, REG_EQUAL, NULL_RTX);
>       if (!note)
>         note = find_reg_note (noted_insn, REG_EQUIV, NULL_RTX);
> +
> +     /* This assert is generally triggered when someone deletes a
> +        REG_EQUAL or REG_EQUIV note by hacking the list manually
> +        rather than calling remove_note.  */
>       gcc_assert (note);
>       remove_note (noted_insn, note);
>     }
> @@ -223,10 +227,34 @@ delete_unmarked_insns (void)
>    something_changed = false;
>    FOR_EACH_BB (bb)
>      FOR_BB_INSNS_SAFE (bb, insn, next)
> -      if (INSN_P (insn)
> -   && ((!marked_insn_p (insn)) || noop_move_p (insn))
> -        && dbg_cnt (new_dce))
> +      if (INSN_P (insn))
>     {
> +     if (noop_move_p (insn))
> +       {
> +         /* Note that this code does not handle the case where
> +       the last insn of libcall is deleted.  As it turns out
> +       this case is excluded in the call to noop_move_p.  */
> +         rtx note = find_reg_note (insn, REG_LIBCALL, NULL_RTX);
> +         if (note && (XEXP (note, 0) != insn))
> +      {
> +        rtx new_libcall_insn = next_real_insn (insn);
> +        rtx retval_note = find_reg_note (XEXP (note, 0),
> +                     REG_RETVAL, NULL_RTX);
> +        REG_NOTES (new_libcall_insn)
> +          = gen_rtx_INSN_LIST (REG_LIBCALL, XEXP (note, 0),
> +                REG_NOTES (new_libcall_insn));
> +        XEXP (retval_note, 0) = new_libcall_insn;
> +      }
> +       }
> +     else if (marked_insn_p (insn))
> +       continue;
> +
> +     /* WARNING, this debugging can itself cause problems if the
> +        edge of the counter causes part of a libcall to be
> +        deleted but not all of it.  */
> +     if (!dbg_cnt (new_dce))
> +       continue;
> +
>       if (dump_file)
>         fprintf (dump_file, "DCE: Deleting insn %d\n", INSN_UID (insn));
>
> @@ -235,8 +263,7 @@ delete_unmarked_insns (void)
>               to prevent dangling REG_EQUAL. */
>            delete_corresponding_reg_eq_notes (insn);
>
> -     /* XXX: This may need to be changed to delete_insn_and_edges */
> -     delete_insn (insn);
> +     delete_insn_and_edges (insn);
>       something_changed = true;
>     }
>  }