Fix to regclass problem

[Richard Kenner]

I just committed the following, which fixes the regclass problem.  The
previous patch was also needed, but this turned out to be the most
important of the two:

Tue Nov  2 21:53:44 1999  Richard Kenner  <kenner@vlsi1.ultra.nyu.edu>

	* regclass.c (record_reg_classes): Always use may_move_cost when
	seeing how operand fits with various register classes.

*** regclass.c	1999/11/01 12:57:38	1.67
--- regclass.c	1999/11/03 02:45:31
*************** record_reg_classes (n_alts, n_ops, ops, 
*** 1250,1254 ****
  		      = (recog_data.operand_type[i] == OP_IN
  			 ? may_move_cost[class][(int) classes[i]]
! 			 : move_cost[(int) classes[i]][class]);
  		  
  		  /* If the alternative actually allows memory, make things
--- 1250,1254 ----
  		      = (recog_data.operand_type[i] == OP_IN
  			 ? may_move_cost[class][(int) classes[i]]
! 			 : may_move_cost[(int) classes[i]][class]);
  		  
  		  /* If the alternative actually allows memory, make things
*************** record_reg_classes (n_alts, n_ops, ops, 
*** 1471,1475 ****
  		      = (recog_data.operand_type[i] == OP_IN
  			 ? may_move_cost[class][(int) classes[i]]
! 			 : move_cost[(int) classes[i]][class]);
  
  		  /* If the alternative actually allows memory, make things
--- 1471,1475 ----
  		      = (recog_data.operand_type[i] == OP_IN
  			 ? may_move_cost[class][(int) classes[i]]
! 			 : may_move_cost[(int) classes[i]][class]);
  
  		  /* If the alternative actually allows memory, make things