> +;; -----------------------------------------------------------------
> +;; ---- Duplicate Operations
> +;; -----------------------------------------------------------------
> +
> +(define_insn_and_split "@vec_duplicate<mode>"
> + [(set (match_operand:VLS 0 "register_operand")
> + (vec_duplicate:VLS
> + (match_operand:<VEL> 1 "reg_or_int_operand")))]
> + "TARGET_VECTOR && can_create_pseudo_p ()"
> + "#"
> + "&& 1"
> + [(const_int 0)]
> + {
> + riscv_vector::emit_vlmax_insn (code_for_pred_broadcast (<MODE>mode),
> + riscv_vector::RVV_UNOP, operands);
> + DONE;
> + }
> +)
Ah, nice. The same thing helps for vv -> vx (for non-VLS modes) because we
can keep the non-expanded version around longer which allows more fwprop
opportunities.
> + /* Enable VECT_COMPARE_COSTS between VLA modes VLS modes for scalable
> + auto-vectorization. */
> + flag |= VECT_COMPARE_COSTS;
I think it's good to activate this here for now. After a while when
costing is reliable we probably want to have it unconditionally even
for pure scalable/fixed-vlmax.
Regards
Robin
