Re: asm constraint rule change

[Bernd Schmidt <crux at pool dot informatik dot rwth-aachen dot de>]

> The recent asm checking patches have tightened the rules such that output
> constraints like "r=" are no longer valid.  The problem is that there is
> a largish body of code that writes things that way.  More to the point,
> despite all the examples running about the documentation, it does not ever
> come out and specify this as a rule.  It merely says 
> 
> 	... all output operands' constraints must use `='.
> 
> So this change is gonna piss some folks off, and with reason.
> 
> I suggest we adopt the following patch to loosen this rule again.
> 
> 	* stmt.c (expand_asm_operands): Accept `=' or `+' at any position.

Then we also need to fix several places in the compiler that assume that
'=' and '+' are always at the first position.  After a quick scan, here are
some examples:

reload.c:
      if (insn_class == NO_REGS
          || (in_p && insn_operand_constraint[(int) icode][!in_p][0] != '=')
          /* The scratch register's constraint must start with "=&".  */
          || insn_operand_constraint[(int) icode][2][0] != '='
          || insn_operand_constraint[(int) icode][2][1] != '&')
        abort ();

local-alloc.c:
          if (1
#ifdef REGISTER_CONSTRAINTS
              && recog_n_operands > 1
              && recog_constraints[0][0] == '='
              && recog_constraints[0][1] != '&'
#else
              && GET_CODE (PATTERN (insn)) == SET
              && rtx_equal_p (SET_DEST (PATTERN (insn)), recog_operand[0])
#endif
              )

Bernd