[Bug debug/60655] [4.9 Regression] ICE: output_operand: invalid expression as operand

[ramrad01 at arm dot com]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60655

--- Comment #6 from ramrad01 at arm dot com ---
On 03/26/14 09:46, jakub at gcc dot gnu.org wrote:
> http://gcc.gnu.org/bugzilla/show_bug.cgi?id=60655
>
> Jakub Jelinek <jakub at gcc dot gnu.org> changed:
>
>             What    |Removed                     |Added
> ----------------------------------------------------------------------------
>                   CC|                            |jakub at gcc dot gnu.org
>
> --- Comment #5 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
> As discussed yesterday with Ramana on IRC, my suggested fix for this for 4.9 is
> something like:
> --- gcc/dwarf2out.c    2014-03-03 08:24:14.841895755 +0100
> +++ gcc/dwarf2out.c    2014-03-26 10:42:32.027508796 +0100
> @@ -11326,7 +11326,12 @@ const_ok_for_output_1 (rtx *rtlp, void *
>       }
>
>     if (GET_CODE (rtl) != SYMBOL_REF)
> -    return 0;
> +    {
> +      /* NOT is not handled by output_addr_const.  */
> +      if (GET_CODE (rtl) == NOT)
> +    return 1;
> +      return 0;
> +    }


I have a very similar patch under testing only with a comment linking 
this to the bug report and bootstrapping and regression testing on 
aarch64, armhf and x86_64.

>
>     if (CONSTANT_POOL_ADDRESS_P (rtl))
>       {
>
> and for 5.0 we want to gather some statistics what we actually accept as CONST
> operands in const_ok_for_output and then successfully emit it into debug info
> and assemble that, and based on that adjust const_ok_for_output or it's
> callers, so that for CONST expressions that have no chance of being assembled
> we actually ignore the CONST around the expression and try to emit it as the
> individual expressions.  So, say instead of trying to emit DW_OP_addr
> (-1-symbol) in this case which won't assemble we emit it as DW_OP_const1s -1
> DW_OP_addr (symbol) DW_OP_minus.

Agreed.


regards
Ramana



>