Re: scratch registers for peephole2

[Denis Chertykov <denisc at overta dot ru>]

Richard Earnshaw <rearnsha@arm.com> writes:

> I've deleted the original mail, but I seem to recall that the AVR port was 
> having problems with scratch registers being used in peephole2 that were 
> not unsafe.  Looking at the code in recog.c, it would appear that there is 
> potentially a similar problem on the ARM with the return register (which 
> is treated as a call-clobbered reg, but can contain real data).

The avr port have this problem only with interrupt or signal
functions.
void foo() __attribute__ ((interrupt));
void foo() 
{
  *vp = v >> 6;
}

In such functions avr port must save/restore all used registers.
`peep2_find_free_register' can choose register from the
CALL_USED_REGISTERS it's normal for ordinar functions, but it's wrong
for AVR "interrupt" and "signal" functions.

We resolve this problem:

(define_peephole2
  [(match_scratch:QI 2 "d")
   (set (match_operand:QI 0 "register_operand" "")
	(match_operand:QI 1 "immediate_operand" ""))]
  "(operands[1] != const0_rtx
    && test_hard_reg_class (NO_LD_REGS, operands[0]))"
  [(parallel [(set (match_dup 0) (match_dup 1))
	      (clobber (match_dup 2))])]
  "if (!avr_peep2_scratch_safe (operands[2]))
     FAIL; /* <--------------- */ ")


int
avr_peep2_scratch_safe (scratch)
     rtx scratch;
{
  if ((interrupt_function_p (current_function_decl)
       || signal_function_p (current_function_decl))
      && leaf_function_p ())
    {
      int first_reg = true_regnum (scratch);
      int last_reg = first_reg + GET_MODE_SIZE (GET_MODE (scratch)) - 1;
      int reg;

      for (reg = first_reg; reg <= last_reg; reg++)
	{
	  if (!regs_ever_live[reg])
	    return 0;
	}
    }
  return 1;
}

> >>>   /* The target may have further reasons for rejecting the register.  
> */
> >>>   if (! REGNO_OK_FOR_PEEPHOLE2 (regno))
> >>>  	continue;
> 
> I'd be inclined to make this macro unconditional, any port using peephole2 
> should define it (it can be 1 if all registers are safe).

Your solution more smart.
Because avr port can reject unsaved registers and agree with the saved
registers if they exists.

Richard Henderson writes:
> Maybe, but your problem on arm with the return register is not,
> repeat NOT here.  It is that you havn't represented that the
> register is live either through rtl or through EPILOGUE_USES.

I don't know anything about arm specific problems. But, for avr your
solution is better than we have. 

Denis.