local-alloc.c (combine_regs): Don't combine if we have a hard reg

[Jim Wilson <wilson at cygnus dot com>]

I have checked in this patch.

This fixes a minor performance problem.  We get an unnecessary move insn
for this example on the SH:

void func(void)
{
  func1(func2());
}

The problem occurs because local-alloc ties the function address of a call
insn with the result register, but this happens to be a single-register
class, and local-alloc isn't supposed to allocate things to single-register
regclasses.  This results in a little confusion later on and we get one more
move insn that we should have gotten.  I have fixed this by disabling tying
if one register is CLASS_LIKELY_SPILLED_P.

It is suspected that we should completely disable tying for CALL_INSNs,
but we don't have any testcase to prove whether this is useful or not.

Mon Mar  8 16:04:44 1999  Jim Wilson  <wilson@cygnus.com>

	* local-alloc.c (combine_regs): Don't combine if we have a hard reg
	for which CLASS_LIKELY_SPILLED_P is true.

Index: local-alloc.c
===================================================================
RCS file: /egcs/carton/cvsfiles/egcs/gcc/local-alloc.c,v
retrieving revision 1.36
diff -p -r1.36 local-alloc.c
*** local-alloc.c	1999/02/25 23:45:25	1.36
--- local-alloc.c	1999/03/09 01:48:39
*************** combine_regs (usedreg, setreg, may_save_
*** 1648,1653 ****
--- 1648,1658 ----
        || ureg == sreg
        /* Don't try to connect two different hardware registers.  */
        || (ureg < FIRST_PSEUDO_REGISTER && sreg < FIRST_PSEUDO_REGISTER)
+       /* Don't use a hard reg that might be spilled.  */
+       || (ureg < FIRST_PSEUDO_REGISTER
+ 	  && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (ureg)))
+       || (sreg < FIRST_PSEUDO_REGISTER
+ 	  && CLASS_LIKELY_SPILLED_P (REGNO_REG_CLASS (sreg)))
        /* Don't connect two different machine modes if they have different
  	 implications as to which registers may be used.  */
        || !MODES_TIEABLE_P (GET_MODE (usedreg), GET_MODE (setreg)))