[PATCH] Generate switch statements in genpreds.c

[Richard Earnshaw]

> 2006-06-27  Roger Sayle  <roger@eyesopen.com>
> 	* genpreds.c (generate_switch_p): New function.
> 	(add_mode_tests): Push the new mode test down inside an AND expr
> 	if this allows the switch-suitable MATCH_CODE to be near the root.
> 	(write_match_code_switch): New function to write a MATCH_CODE as
> 	a switch statement.
> 	(write_predicate_stmts): New function to write a predicate RTX
> 	expression as a sequence of statements.
> 	(write_one_predicate_function): Use write_predicate_stmts.
> 	(write_tm_constrs_h): Likewise.

Unfortunately, on an ARM-eabi cross:

(define_predicate "di_operand"
  (ior (match_code "const_int,const_double")
       (and (match_code "reg,subreg,mem")
            (match_operand 0 "nonimmediate_di_operand"))))

is now compiled to:

di_operand (rtx op, enum machine_mode mode ATTRIBUTE_UNUSED)
{
  switch (GET_CODE (op))
    {
    case CONST_INT:
    case CONST_DOUBLE:
      return true;
    default:
      break;
    }
  switch (GET_CODE (op))
    {
    case REG:
    case SUBREG:
    case MEM:
      return false;
    default:
      break;
    }
  return true;
}

Which is not a correct expansion of the logic of the predicate...  What 
happened to the call to nonimmediate_di_operand?  And why does it return 
true if op is not one of CONST_INT, CONST_DOUBLE, REG, SUBREG or MEM?

R.