stormy16.c

Mac OS X 10.4.9 for x86 Source Code gcc 实现源代码

int
nonimmediate_nonstack_operand (rtx op, enum machine_mode mode)
{
  /* 'Q' is for pushes, 'R' for pops.  */
  return (nonimmediate_operand (op, mode) 
	  && ! xstormy16_extra_constraint_p (op, 'Q')
	  && ! xstormy16_extra_constraint_p (op, 'R'));
}

/* Splitter for the 'move' patterns, for modes not directly implemented
   by hardware.  Emit insns to copy a value of mode MODE from SRC to
   DEST.

   This function is only called when reload_completed.
   */

void 
xstormy16_split_move (enum machine_mode mode, rtx dest, rtx src)
{
  int num_words = GET_MODE_BITSIZE (mode) / BITS_PER_WORD;
  int direction, end, i;
  int src_modifies = 0;
  int dest_modifies = 0;
  int src_volatile = 0;
  int dest_volatile = 0;
  rtx mem_operand;
  rtx auto_inc_reg_rtx = NULL_RTX;
  
  /* Check initial conditions.  */
  if (! reload_completed
      || mode == QImode || mode == HImode
      || ! nonimmediate_operand (dest, mode)
      || ! general_operand (src, mode))
    abort ();

  /* This case is not supported below, and shouldn't be generated.  */
  if (GET_CODE (dest) == MEM
      && GET_CODE (src) == MEM)
    abort ();

  /* This case is very very bad after reload, so trap it now.  */
  if (GET_CODE (dest) == SUBREG
      || GET_CODE (src) == SUBREG)
    abort ();

  /* The general idea is to copy by words, offsetting the source and
     destination.  Normally the least-significant word will be copied
     first, but for pre-dec operations it's better to copy the 
     most-significant word first.  Only one operand can be a pre-dec
     or post-inc operand.  

     It's also possible that the copy overlaps so that the direction
     must be reversed.  */
  direction = 1;
  
  if (GET_CODE (dest) == MEM)
    {
      mem_operand = XEXP (dest, 0);
      dest_modifies = side_effects_p (mem_operand);
      if (auto_inc_p (mem_operand))
        auto_inc_reg_rtx = XEXP (mem_operand, 0);
      dest_volatile = MEM_VOLATILE_P (dest);
      if (dest_volatile)
	{
	  dest = copy_rtx (dest);
	  MEM_VOLATILE_P (dest) = 0;
	}
    }
  else if (GET_CODE (src) == MEM)
    {
      mem_operand = XEXP (src, 0);
      src_modifies = side_effects_p (mem_operand);
      if (auto_inc_p (mem_operand))
        auto_inc_reg_rtx = XEXP (mem_operand, 0);
      src_volatile = MEM_VOLATILE_P (src);
      if (src_volatile)
	{
	  src = copy_rtx (src);
	  MEM_VOLATILE_P (src) = 0;
	}
    }
  else
    mem_operand = NULL_RTX;

  if (mem_operand == NULL_RTX)
    {
      if (GET_CODE (src) == REG
	  && GET_CODE (dest) == REG
	  && reg_overlap_mentioned_p (dest, src)
	  && REGNO (dest) > REGNO (src))
	direction = -1;
    }
  else if (GET_CODE (mem_operand) == PRE_DEC
      || (GET_CODE (mem_operand) == PLUS 
	  && GET_CODE (XEXP (mem_operand, 0)) == PRE_DEC))
    direction = -1;
  else if (GET_CODE (src) == MEM
	   && reg_overlap_mentioned_p (dest, src))
    {
      int regno;
      if (GET_CODE (dest) != REG)
	abort ();
      regno = REGNO (dest);
      
      if (! refers_to_regno_p (regno, regno + num_words, mem_operand, 0))
	abort ();
      
      if (refers_to_regno_p (regno, regno + 1, mem_operand, 0))
	direction = -1;
      else if (refers_to_regno_p (regno + num_words - 1, regno + num_words,
				  mem_operand, 0))
	direction = 1;
      else
	/* This means something like
	   (set (reg:DI r0) (mem:DI (reg:HI r1)))
	   which we'd need to support by doing the set of the second word
	   last.  */
	abort ();
    }

  end = direction < 0 ? -1 : num_words;
  for (i = direction < 0 ? num_words - 1 : 0; i != end; i += direction)
    {
      rtx w_src, w_dest, insn;

      if (src_modifies)
	w_src = gen_rtx_MEM (word_mode, mem_operand);
      else
	w_src = simplify_gen_subreg (word_mode, src, mode, i * UNITS_PER_WORD);
      if (src_volatile)
	MEM_VOLATILE_P (w_src) = 1;
      if (dest_modifies)
	w_dest = gen_rtx_MEM (word_mode, mem_operand);
      else
	w_dest = simplify_gen_subreg (word_mode, dest, mode, 
				      i * UNITS_PER_WORD);
      if (dest_volatile)
	MEM_VOLATILE_P (w_dest) = 1;
      
      /* The simplify_subreg calls must always be able to simplify.  */
      if (GET_CODE (w_src) == SUBREG
	  || GET_CODE (w_dest) == SUBREG)
	abort ();
      
      insn = emit_insn (gen_rtx_SET (VOIDmode, w_dest, w_src));
      if (auto_inc_reg_rtx)
        REG_NOTES (insn) = alloc_EXPR_LIST (REG_INC,
                                            auto_inc_reg_rtx,
					    REG_NOTES (insn));
    }
}

/* Expander for the 'move' patterns.  Emit insns to copy a value of
   mode MODE from SRC to DEST.  */

void 
xstormy16_expand_move (enum machine_mode mode, rtx dest, rtx src)
{
  if ((GET_CODE (dest) == MEM) && (GET_CODE (XEXP (dest, 0)) == PRE_MODIFY))
    {
      rtx pmv      = XEXP (dest, 0);
      rtx dest_reg = XEXP (pmv, 0);
      rtx dest_mod = XEXP (pmv, 1);
      rtx set      = gen_rtx_SET (Pmode, dest_reg, dest_mod);
      rtx clobber  = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (BImode, 16));
    
      dest = gen_rtx_MEM (mode, dest_reg);
      emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, set, clobber)));
    }
  else if ((GET_CODE (src) == MEM) && (GET_CODE (XEXP (src, 0)) == PRE_MODIFY))
    {
      rtx pmv     = XEXP (src, 0);
      rtx src_reg = XEXP (pmv, 0);
      rtx src_mod = XEXP (pmv, 1);
      rtx set     = gen_rtx_SET (Pmode, src_reg, src_mod);
      rtx clobber = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (BImode, 16));
    
      src = gen_rtx_MEM (mode, src_reg);
      emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, set, clobber)));
    }
   
  /* There are only limited immediate-to-memory move instructions.  */
  if (! reload_in_progress
      && ! reload_completed
      && GET_CODE (dest) == MEM
      && (GET_CODE (XEXP (dest, 0)) != CONST_INT
	  || ! xstormy16_legitimate_address_p (mode, XEXP (dest, 0), 0))
      && ! xstormy16_below100_operand (dest, mode)
      && GET_CODE (src) != REG
      && GET_CODE (src) != SUBREG)
    src = copy_to_mode_reg (mode, src);

  /* Don't emit something we would immediately split.  */
  if (reload_completed
      && mode != HImode && mode != QImode)
    {
      xstormy16_split_move (mode, dest, src);
      return;
    }
  
  emit_insn (gen_rtx_SET (VOIDmode, dest, src));
}


/* Stack Layout:

   The stack is laid out as follows:

SP->
FP->	Local variables
	Register save area (up to 4 words)
	Argument register save area for stdarg (NUM_ARGUMENT_REGISTERS words)

AP->	Return address (two words)
	9th procedure parameter word
	10th procedure parameter word
	...
	last procedure parameter word

  The frame pointer location is tuned to make it most likely that all
  parameters and local variables can be accessed using a load-indexed
  instruction.  */

/* A structure to describe the layout.  */
struct xstormy16_stack_layout
{
  /* Size of the topmost three items on the stack.  */
  int locals_size;
  int register_save_size;
  int stdarg_save_size;
  /* Sum of the above items.  */
  int frame_size;
  /* Various offsets.  */
  int first_local_minus_ap;
  int sp_minus_fp;
  int fp_minus_ap;
};

/* Does REGNO need to be saved?  */
#define REG_NEEDS_SAVE(REGNUM, IFUN)					\
  ((regs_ever_live[REGNUM] && ! call_used_regs[REGNUM])			\
   || (IFUN && ! fixed_regs[REGNUM] && call_used_regs[REGNUM]		\
       && (REGNO_REG_CLASS (REGNUM) != CARRY_REGS)			\
       && (regs_ever_live[REGNUM] || ! current_function_is_leaf)))

/* Compute the stack layout.  */
struct xstormy16_stack_layout 
xstormy16_compute_stack_layout (void)
{
  struct xstormy16_stack_layout layout;
  int regno;
  const int ifun = xstormy16_interrupt_function_p ();

  layout.locals_size = get_frame_size ();
  
  layout.register_save_size = 0;
  for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
    if (REG_NEEDS_SAVE (regno, ifun))
      layout.register_save_size += UNITS_PER_WORD;
  
  if (current_function_stdarg)
    layout.stdarg_save_size = NUM_ARGUMENT_REGISTERS * UNITS_PER_WORD;
  else
    layout.stdarg_save_size = 0;
  
  layout.frame_size = (layout.locals_size 
		       + layout.register_save_size 
		       + layout.stdarg_save_size);
  
  if (current_function_args_size <= 2048 && current_function_args_size != -1)
    {
      if (layout.frame_size + INCOMING_FRAME_SP_OFFSET 
	  + current_function_args_size <= 2048)
	layout.fp_minus_ap = layout.frame_size + INCOMING_FRAME_SP_OFFSET;
      else
	layout.fp_minus_ap = 2048 - current_function_args_size;
    }
  else
    layout.fp_minus_ap = (layout.stdarg_save_size 
			  + layout.register_save_size
			  + INCOMING_FRAME_SP_OFFSET);
  layout.sp_minus_fp = (layout.frame_size + INCOMING_FRAME_SP_OFFSET 
			- layout.fp_minus_ap);
  layout.first_local_minus_ap = layout.sp_minus_fp - layout.locals_size;
  return layout;
}

/* Determine how all the special registers get eliminated.  */
int
xstormy16_initial_elimination_offset (int from, int to)
{
  struct xstormy16_stack_layout layout;
  int result;
  
  layout = xstormy16_compute_stack_layout ();

  if (from == FRAME_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
    result = layout.sp_minus_fp - layout.locals_size;
  else if (from == FRAME_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
    result = -layout.locals_size;
  else if (from == ARG_POINTER_REGNUM && to == HARD_FRAME_POINTER_REGNUM)
    result = -layout.fp_minus_ap;
  else if (from == ARG_POINTER_REGNUM && to == STACK_POINTER_REGNUM)
    result = -(layout.sp_minus_fp + layout.fp_minus_ap);
  else
    abort ();

  return result;
}

static rtx
emit_addhi3_postreload (rtx dest, rtx src0, rtx src1)
{
  rtx set, clobber, insn;
  
  set = gen_rtx_SET (VOIDmode, dest, gen_rtx_PLUS (HImode, src0, src1));
  clobber = gen_rtx_CLOBBER (VOIDmode, gen_rtx_REG (BImode, 16));
  insn = emit_insn (gen_rtx_PARALLEL (VOIDmode, gen_rtvec (2, set, clobber)));
  return insn;
}

/* Called after register allocation to add any instructions needed for
   the prologue.  Using a prologue insn is favored compared to putting
   all of the instructions in the TARGET_ASM_FUNCTION_PROLOGUE macro,
   since it allows the scheduler to intermix instructions with the
   saves of the caller saved registers.  In some cases, it might be
   necessary to emit a barrier instruction as the last insn to prevent
   such scheduling.

   Also any insns generated here should have RTX_FRAME_RELATED_P(insn) = 1
   so that the debug info generation code can handle them properly.  */
void
xstormy16_expand_prologue (void)
{
  struct xstormy16_stack_layout layout;
  int regno;
  rtx insn;
  rtx mem_push_rtx;
  const int ifun = xstormy16_interrupt_function_p ();
  
  mem_push_rtx = gen_rtx_POST_INC (Pmode, stack_pointer_rtx);
  mem_push_rtx = gen_rtx_MEM (HImode, mem_push_rtx);
    
  layout = xstormy16_compute_stack_layout ();

  if (layout.locals_size >= 32768)
    error ("Local variable memory requirements exceed capacity.");

  /* Save the argument registers if necessary.  */
  if (layout.stdarg_save_size)
    for (regno = FIRST_ARGUMENT_REGISTER; 
	 regno < FIRST_ARGUMENT_REGISTER + NUM_ARGUMENT_REGISTERS;
	 regno++)
      {
	rtx dwarf;
	rtx reg = gen_rtx_REG (HImode, regno);

	insn = emit_move_insn (mem_push_rtx, reg);
	RTX_FRAME_RELATED_P (insn) = 1;

	dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (2));
	
	XVECEXP (dwarf, 0, 0) = gen_rtx_SET (VOIDmode,
					     gen_rtx_MEM (Pmode, stack_pointer_rtx),
					     reg);
	XVECEXP (dwarf, 0, 1) = gen_rtx_SET (Pmode, stack_pointer_rtx,
					     plus_constant (stack_pointer_rtx,
							    GET_MODE_SIZE (Pmode)));
	REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
					      dwarf,
					      REG_NOTES (insn));
	RTX_FRAME_RELATED_P (XVECEXP (dwarf, 0, 0)) = 1;
	RTX_FRAME_RELATED_P (XVECEXP (dwarf, 0, 1)) = 1;
      }
  
  /* Push each of the registers to save.  */
  for (regno = 0; regno < FIRST_PSEUDO_REGISTER; regno++)
    if (REG_NEEDS_SAVE (regno, ifun))
      {
	rtx dwarf;
	rtx reg = gen_rtx_REG (HImode, regno);

	insn = emit_move_insn (mem_push_rtx, reg);
	RTX_FRAME_RELATED_P (insn) = 1;

	dwarf = gen_rtx_SEQUENCE (VOIDmode, rtvec_alloc (2));
	
	XVECEXP (dwarf, 0, 0) = gen_rtx_SET (VOIDmode,
					     gen_rtx_MEM (Pmode, stack_pointer_rtx),
					     reg);
	XVECEXP (dwarf, 0, 1) = gen_rtx_SET (Pmode, stack_pointer_rtx,
					     plus_constant (stack_pointer_rtx,
							    GET_MODE_SIZE (Pmode)));
	REG_NOTES (insn) = gen_rtx_EXPR_LIST (REG_FRAME_RELATED_EXPR,
					      dwarf,
					      REG_NOTES (insn));
	RTX_FRAME_RELATED_P (XVECEXP (dwarf, 0, 0)) = 1;
	RTX_FRAME_RELATED_P (XVECEXP (dwarf, 0, 1)) = 1;
      }

  /* It's just possible that the SP here might be what we need for
     the new FP...  */
  if (frame_pointer_needed && layout.sp_minus_fp == layout.locals_size)
    emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx);

  /* Allocate space for local variables.  */
  if (layout.locals_size)
    {
      insn = emit_addhi3_postreload (stack_pointer_rtx, stack_pointer_rtx,
				     GEN_INT (layout.locals_size));
      RTX_FRAME_RELATED_P (insn) = 1;
    }

  /* Set up the frame pointer, if required.  */
  if (frame_pointer_needed && layout.sp_minus_fp != layout.locals_size)
    {
      insn = emit_move_insn (hard_frame_pointer_rtx, stack_pointer_rtx);

      if (layout.sp_minus_fp)
	emit_addhi3_postreload (hard_frame_pointer_rtx,
				hard_frame_pointer_rtx,
				GEN_INT (-layout.sp_minus_fp));
    }
}

/* Do we need an epilogue at all?  */
int
direct_return (void)
{
  return (reload_completed 
	  && xstormy16_compute_stack_layout ().frame_size == 0);
}

/* Called after register allocation to add any instructions needed for
   the epilogue.  Using an epilogue insn is favored compared to putting
   all of the instructions in the TARGET_ASM_FUNCTION_PROLOGUE macro,
   since it allows the scheduler to intermix instructions with the