Bogus rtx generated?

[Toshiyasu Morita]

>   In message < 199803161156.DAA00822@netcom9.netcom.com >you write:
>   > 
>   > I'm tracking down a bug which causes gcc to crash; it looks like reload is
>   > crashing when trying to generate reloads for the following rtx:
>   > 
>   > (insn 349 33 350 (parallel[
>   >             (set (subreg:SI (reg/v:DI 46) 0)
>   >                 (ashift:SI (subreg:SI (reg/v:DI 44) 0)
>   >                     (const_int 1)))
>   >             (set (reg:SI 18 t)
>   >                 (lt:SI (subreg:SI (reg/v:DI 44) 0)
>   >                     (const_int 0)))
>   >         ] ) -1 (nil)
>   >     (nil))
>   > 
>   > Is this a reasonable rtx? I don't understand why you would have a parallel
>   > shift with a compare, so this seems rather fishy to me.
> There's nothing wrong with it if the machine supports this kind of
> instruction.  I don't know what target you're using and thus I can't
> hazard a guess as to whether or not this is a suitable instruction
> for your machine.

It's the Hitachi SH. AFAIK it doesn't have this instruction.

>   > It looks like the previous rtx is generated by the first scheduling pass;
>   > here's the rtx before scheduling:
>   > 
>   > (insn 35 33 36 (parallel[
>   >             (set (subreg:SI (reg/v:DI 46) 0)
>   >                 (ashiftrt:SI (subreg:SI (reg/v:DI 44) 0)
>   >                     (const_int 31)))
>   >             (clobber (reg:SI 18 t))
>   >         ] ) 68 {ashrsi2_31} (nil)
>   >     (expr_list:REG_UNUSED (reg:SI 18 t)
>   >         (nil)))
>   > 
>   > but after try_split() is called, the first rtx is generated...
> This is certainly possible if the target has a define_spit which
> isn't operating correctly.

It looks like this split matches the rtx - am I correct?

(define_split
  [(set (match_operand:SI 0 "arith_reg_operand" "=r")
        (ashiftrt:SI (match_operand:SI 1 "arith_reg_operand" "0")
                     (const_int 31)))
   (clobber (reg:SI 18))]
  ""
  [(const_int 0)]
  "
{
  emit_insn (gen_ashlsi_c (operands[0], operands[1], operands[1]));
  emit_insn (gen_subc1 (operands[0], operands[0], operands[0]));
  DONE;
}")

> jeff

Toshi