static Temp emit_bpermute(isel_context *ctx, Builder &bld, Definition dst, Temp index, Temp data)
{
   if (ctx->options->chip_class <= GFX9) {
      return bld.ds(aco_opcode::ds_bpermute_b32, dst, index, data);
   }

   /* XXX: only do this for wave64 */

   Temp swapped_data = bld.pseudo(aco_opcode::p_swap_half_waves, bld.def(v1), data, Operand(s1));

   Temp lane_id = bld.vop3(aco_opcode::v_mbcnt_lo_u32_b32, bld.def(v1), Operand((uint32_t) -1), Operand(0u));
   lane_id = bld.vop3(aco_opcode::v_mbcnt_hi_u32_b32, bld.def(v1), Operand((uint32_t) -1), lane_id);
   Temp lane_is_hi = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), lane_id, Operand(0x20u));
   Temp index_is_hi = bld.vop2(aco_opcode::v_and_b32, bld.def(v1), index, Operand(0x20u));

   Temp bperm_1 = bld.ds(aco_opcode::ds_bpermute_b32, bld.def(v1), index, data);
   Temp bperm_2 = bld.ds(aco_opcode::ds_bpermute_b32, bld.def(v1), index, swapped_data);

   Definition vcc_def = bld.def(v1);
   vcc_def.setHint(vcc);
   Temp vcc_tmp = bld.vopc(aco_opcode::v_cmp_eq_u32, vcc_def, lane_is_hi, index_is_hi);
   return bld.vop2(aco_opcode::v_cndmask_b32, dst, bperm_2, bperm_1, vcc_tmp);
}