sh-elf: PR target/15886

[Kaz Kojima <kkojima at rr dot iij4u dot or dot jp>]

Hi,

I could reduce the testcase of PR target/15886 to

register int r8 asm ("r8");
register int r9 asm ("r9");
register int r10 asm ("r10");
register int r11 asm ("r11");
register int r12 asm ("r12");
register int r13 asm ("r13");

void test (void)
{
  void *temp;
  int *arg[4];

  dummy ();
  temp = __builtin_return_address (0);
  arg[0] = 0;
  arg[1] = temp;
  foo (arg);
}

The mainline and 3.4 sh-elf gcc assembles it to

        mov.l   r14,@-r15
        sts.l   pr,@-r15
        add     #-16,r15
        mov     r15,r14
        sts     pr,r1
        mov.l   .L2,r0
        jsr     @r0
        mov.l   r1,@r14
        mov     #0,r1
        mov.l   r1,@r14
        mov.l   r1,@(4,r14)
        mov.l   .L3,r0
        jsr     @r0
        mov     r14,r4
        add     #16,r14
        mov     r14,r15
        lds.l   @r15+,pr
        rts

with -O1.

It seems that pr register is saved to a register allocated in
stack and reload prefers @r14 for that register, while @r14
is allocated for arg[0].  I've found that reg_equiv_memory_loc[i]
for that register is set to @r14 in allocate_initial_values
with ALLOCATE_INITIAL_VALUE macro and sh defines it like as

#define ALLOCATE_INITIAL_VALUE(hard_reg) \
  (REGNO (hard_reg) == (TARGET_SHMEDIA ? PR_MEDIA_REG : PR_REG) \
   ? (current_function_is_leaf \
      && ! sh_pr_n_sets () \
      && ! (TARGET_SHCOMPACT \
            && ((current_function_args_info.call_cookie \
                 & ~ CALL_COOKIE_RET_TRAMP (1)) \
                || current_function_has_nonlocal_label)) \
      ? (hard_reg) \
      : gen_rtx_MEM (Pmode, TARGET_SH5 \
                            ? (plus_constant (arg_pointer_rtx, \
                                              TARGET_SHMEDIA64 ? -8 : -4)) \
                            : frame_pointer_rtx)) \
   : NULL_RTX)

This looks odd for TARGET_SH1.  pr is saved at the address pointed
by frame_pointer_rtx + frame size for TARGET_SH1, but it seems that
the final frame size can't be known here because allocate_initial_values
is called in regalloc phase.
Is it wrong to just return NULL_RTX for that case?

Regards,
	kaz