minor bug (?) in alpha.c

[kanazawa at flab dot fujitsu dot co dot jp (Kanazawa Yuzi)]

  I tried egcs-971031 for alpha-linux. I am happy that code produced
by egcs with haifa scheduler runs 10 - 20% faster than code produced
by gcc-2.7.2.1.  Yet I think there is a minor bug in the code added to
use haifa scheduler.

  The following is the code in alpha_adjust_cost() which exists after
line 1193 of gcc/config/alpha/alpha.c . I think this code is strange
because it never checks attr_type of `dep_insn', but checks attr_type
of `insn' twice. I guess this is a typo and the argument of switch
statement should be `get_attr_type(dep_insn)'.

----- cut from gcc/config/alpha/alpha.c -----
      /* On EV5 it takes longer to get data to the multiplier than to
	 anywhere else, so increase costs.  */

      if (recog_memoized (insn) >= 0
	  && recog_memoized (dep_insn) >= 0
	  && (get_attr_type (insn) == TYPE_IMULL
	      || get_attr_type (insn) == TYPE_IMULQ
	      || get_attr_type (insn) == TYPE_IMULH)
	  && (set = single_set (dep_insn)) != 0
	  && GET_CODE (PATTERN (insn)) == SET
	  && (set_src = SET_SRC (PATTERN (insn)),
	      GET_CODE (set_src) == MULT)
	  && (set = SET_DEST (set),
	      rtx_equal_p (set, XEXP (set_src, 0))
	      || rtx_equal_p (set, XEXP (set_src, 1))))
	{
	  switch (get_attr_type (insn))
	    {
	    case TYPE_LD:
	    case TYPE_CMOV:
	    case TYPE_IMULL:
	    case TYPE_IMULQ:
	    case TYPE_IMULH:
	      return cost + 1;
	    case TYPE_JSR:
	    case TYPE_IADD:
	    case TYPE_ILOG:
	    case TYPE_SHIFT:
	    case TYPE_ICMP:
	      return cost + 2;
	    }
	}

-- 
Kanazawa, Yuzi
kanazawa@flab.fujitsu.co.jp