[Bug target/37489] In cse.c:fold_rtx(), "true" is represented in floating-point modes as const_true_rtx, if FLOAT_STORE_FLAG_VALUE is undefined.

["raksit at gcc dot gnu dot org" <gcc-bugzilla@gcc.gnu.org>]

------- Comment #6 from raksit at gcc dot gnu dot org  2008-09-12 18:54 -------
(In reply to comment #4)
> We can't define FLOAT_STORE_FLAG_VALUE for SSE since we can't
> represent 0xfffffff as a valid FP value. This patch makes
> fold_rtx to match simplify_relational_operation:
> 
> --- ./cse.c.foo 2008-09-08 10:46:09.000000000 -0700
> +++ ./cse.c     2008-09-12 10:46:18.000000000 -0700
> @@ -3217,14 +3217,16 @@ fold_rtx (rtx x, rtx insn)
>           rtx true_rtx = const_true_rtx, false_rtx = const0_rtx;
>           enum machine_mode mode_arg1;
> 
> -#ifdef FLOAT_STORE_FLAG_VALUE
>           if (SCALAR_FLOAT_MODE_P (mode))
>             {
> +#ifdef FLOAT_STORE_FLAG_VALUE
>               true_rtx = (CONST_DOUBLE_FROM_REAL_VALUE
>                           (FLOAT_STORE_FLAG_VALUE (mode), mode));
> +#else
> +             true_rtx = NULL_RTX;
> +#endif
>               false_rtx = CONST0_RTX (mode);
>             }
> -#endif
> 
>           code = find_comparison_args (code, &folded_arg0, &folded_arg1,
>                                        &mode_arg0, &mode_arg1);
> @@ -3332,8 +3334,17 @@ fold_rtx (rtx x, rtx insn)
>                                                   const_arg1))
>                               || (REG_P (folded_arg1)
>                                   && (REG_QTY (REGNO (folded_arg1)) ==
> ent->comparison_qty))))
> -                       return (comparison_dominates_p (ent->comparison_code,
> code)
> -                               ? true_rtx : false_rtx);
> +                       {
> +                         if (comparison_dominates_p (ent->comparison_code,
> code))
> +                           {
> +                             if (true_rtx)
> +                               return true_rtx;
> +                             else
> +                               break;
> +                           }
> +                         else
> +                           return false_rtx;
> +                       }
>                     }
>                 }
>             }
> 

I had a similar patch in mind. In case it helps, I have attached a testcase
that you can use for the submission. You might also want to fix the gcc-4.3
branch.

-raksit


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=37489