optimization/10592: [arm] caller-save incorrectly handles a reloaded SET_DEST operand

[Richard Earnshaw]

[Jeff, I've put you on CC because you appear to be the original author of 
the relevant code (back in Oct 98).]

Some more detailed analysis...

The problem we have here comes from the use of caller-save register 
allocation and the failure to allocate a register to a pseudo that is 
equivalent to a memory location that is derived from a caller-saved 
register value.  Specifically, we have the sequence prior to reload of

(call_insn:HI 249 227 326 2 0x4067c5d8 (parallel [
            (set (reg:SI 0 r0)
                (call (mem:SI (symbol_ref:SI ("strncpy") <function_decl 
0x40330af8 strncpy>) [0 S4 A32])
                    (const_int 0 [0x0])))
            (use (const_int 0 [0x0]))
            (clobber (reg:SI 14 lr))
        ]) 190 {*call_value_symbol} (insn_list:REG_DEP_ANTI 76 (insn_list 
248 (insn_list 247 (insn_list 246 (insn_list:REG_DEP_ANTI 30 
(insn_list:REG_DEP_ANTI 172 (insn_list 210 (insn_list:REG_DEP_ANTI 75 
(insn_list:REG_DEP_ANTI 74 (insn_list 47 (insn_list 227 (nil))))))))))))
    (expr_list:REG_DEAD (reg:SI 1 r1 [ pName ])
        (expr_list:REG_DEAD (reg:SI 2 r2)
            (expr_list:REG_UNUSED (reg:SI 0 r0)
                (expr_list:REG_UNUSED (reg:SI 14 lr)
                    (expr_list:REG_EH_REGION (const_int 0 [0x0])
                        (nil))))))
    (expr_list (use (reg:SI 2 r2))
        (expr_list (use (reg:SI 1 r1 [ pName ]))
            (expr_list (use (reg:SI 0 r0))
                (nil)))))

(insn:HI 326 249 331 2 0x40658dc0 (set (reg:SI 98)
        (plus:SI (reg/v:SI 49 [ base ])
            (const_int 156 [0x9c]))) 4 {*arm_addsi3} (insn_list 4 
(insn_list:REG_DEP_ANTI 249 (nil)))
    (expr_list:REG_EQUIV (mem/s:SI (plus:SI (reg/v/u/f:SI 48 [ this ])
                (const_int 540 [0x21c])) [39 <variable>.pValue3+0 S4 A32])
        (nil)))

Note that pseudo 98 has a REG_EQUIV of mem(plus (r48, 540)) 

Now global-alloc assigns r48 to the call-clobbered register r12, but 
spills r98 back into its reg_equiv location.  When save_call_clobbered_res 
runs, therefore, r12 is stored to the stack before the call insn, but when 
scanning r326 mark_referenced_regs ignores the SET_DEST in insn 326 
because it is a REG:

  if (code == SET || code == CLOBBER)
    {
      x = SET_DEST (x);
      code = GET_CODE (x);
      if (code == REG || code == PC || code == CC0
          || (code == SUBREG && GET_CODE (SUBREG_REG (x)) == REG
              /* If we're setting only part of a multi-word register,
                 we shall mark it as referenced, because the words
                 that are not being set should be restored.  */
              && ((GET_MODE_SIZE (GET_MODE (x))
                   >= GET_MODE_SIZE (GET_MODE (SUBREG_REG (x))))
                  || (GET_MODE_SIZE (GET_MODE (SUBREG_REG (x)))
                      <= UNITS_PER_WORD))))
        return;

We thus end up with the sequence of instructions

	str	ip, [frame+xxx]
	call	strncpy
	add	tmp, r9, #156
	str	tmp, [ip+540]	// ip not restored first


I think the fix may be to make mark_referenced_regs only return if REG is 
a hard register (similarly for SUBREG).

Jeff, any thoughts?

R.