On 26/11/13 10:07, Charles Baylis wrote:
> Updated version. Only build tested.
> 
>         PR target/59142
>         gcc/
>         * arm/predicates.md (arm_hard_general_register_operand): New 
> predicate.
>         (vfp_hard_register_operand): Likewise.
>         * config/arm/arm-ldmstm.ml: Use
> arm_hard_general_register_operand for all patterns.
>         * config/arm/ldmstm.md: Regenerate.
>         * gcc/config/arm/arm.md: (vfp_pop_multiple_with_writeback) Use
> vfp_hard_register_operand.
> 

Watch out for over-long lines (79 column limit).

> On 25 November 2013 14:54, Richard Earnshaw <rearn...@arm.com> wrote:
>> You should delete the predicate arm_hard_register_operand if it's now
>> no-longer being used.  Don't forget the ChangeLog entry for that as well.
> 
> I didn't change uses that aren't inside a define_insn, so the old
> predicate is still used for now.
> 

Grep-ing the sources doesn't show any other obvious uses outside of
define_insn; so what's left?

I've just spotted another problem (that was always there):
> @@ -42,15 +42,15 @@
>  
>  (define_insn "*thumb_ldm4_ia"
>    [(match_parallel 0 "load_multiple_operation"
> -    [(set (match_operand:SI 1 "arm_hard_register_operand" "")
> +    [(set (match_operand:SI 1 "arm_hard_general_register_operand" "")
>            (mem:SI (match_operand:SI 5 "s_register_operand" "l")))
> -     (set (match_operand:SI 2 "arm_hard_register_operand" "")
> +     (set (match_operand:SI 2 "arm_hard_general_register_operand" "")
>            (mem:SI (plus:SI (match_dup 5)
>                    (const_int 4))))
> -     (set (match_operand:SI 3 "arm_hard_register_operand" "")
> +     (set (match_operand:SI 3 "arm_hard_general_register_operand" "")
>            (mem:SI (plus:SI (match_dup 5)
>                    (const_int 8))))
> -     (set (match_operand:SI 4 "arm_hard_register_operand" "")
> +     (set (match_operand:SI 4 "arm_hard_general_register_operand" "")
>            (mem:SI (plus:SI (match_dup 5)
>                    (const_int 12))))])]
>    "TARGET_THUMB1 && XVECLEN (operands[0], 0) == 4"

This, and other thumb1 patterns should be stricter than this, since the
T1 LDM/STM patterns only support low regs.  I think these need changing
to low_register_operand.

R.


Reply via email to