m88k.c

linux下的gcc编译器

#define MOVSTR_DI	96 /* __movstrDI96x96 .. __movstrDI96x16 */
#define MOVSTR_ODD_HI	16 /* __movstrHI15x15 .. __movstrHI15x5 */
#define MOVSTR_ODD_SI	48 /* __movstrSI47x47 .. __movstrSI47x11,
			      __movstrSI46x46 .. __movstrSI46x10,
			      __movstrSI45x45 .. __movstrSI45x9 */
#define MOVSTR_ODD_DI	48 /* __movstrDI47x47 .. __movstrDI47x23,
			      __movstrDI46x46 .. __movstrDI46x22,
			      __movstrDI45x45 .. __movstrDI45x21,
			      __movstrDI44x44 .. __movstrDI44x20,
			      __movstrDI43x43 .. __movstrDI43x19,
			      __movstrDI42x42 .. __movstrDI42x18,
			      __movstrDI41x41 .. __movstrDI41x17 */

/* Limits for using the non-looping movstr functions.  For the m88100
   processor, we assume the source and destination are word aligned.
   The QImode and HImode limits are the break even points where memcpy
   does just as well and beyond which memcpy does better.  For the
   m88110, we tend to assume double word alignment, but also analyze
   the word aligned cases.  The analysis is complicated because memcpy
   may use the cache control instructions for better performance.  */

#define MOVSTR_QI_LIMIT_88100   13
#define MOVSTR_HI_LIMIT_88100   38
#define MOVSTR_SI_LIMIT_88100   MOVSTR_SI
#define MOVSTR_DI_LIMIT_88100   MOVSTR_SI
  
#define MOVSTR_QI_LIMIT_88000   16
#define MOVSTR_HI_LIMIT_88000   38
#define MOVSTR_SI_LIMIT_88000   72
#define MOVSTR_DI_LIMIT_88000   72

#define MOVSTR_QI_LIMIT_88110   16
#define MOVSTR_HI_LIMIT_88110   38
#define MOVSTR_SI_LIMIT_88110   72
#define MOVSTR_DI_LIMIT_88110   72

static const enum machine_mode mode_from_align[] =
			      {VOIDmode, QImode, HImode, VOIDmode, SImode,
			       VOIDmode, VOIDmode, VOIDmode, DImode};
static const int max_from_align[] = {0, MOVSTR_QI, MOVSTR_HI, 0, MOVSTR_SI,
				     0, 0, 0, MOVSTR_DI};
static const int all_from_align[] = {0, MOVSTR_QI, MOVSTR_ODD_HI, 0,
				     MOVSTR_ODD_SI, 0, 0, 0, MOVSTR_ODD_DI};

static const int best_from_align[3][9] = {
  {0, MOVSTR_QI_LIMIT_88100, MOVSTR_HI_LIMIT_88100, 0, MOVSTR_SI_LIMIT_88100,
   0, 0, 0, MOVSTR_DI_LIMIT_88100},
  {0, MOVSTR_QI_LIMIT_88110, MOVSTR_HI_LIMIT_88110, 0, MOVSTR_SI_LIMIT_88110,
   0, 0, 0, MOVSTR_DI_LIMIT_88110},
  {0, MOVSTR_QI_LIMIT_88000, MOVSTR_HI_LIMIT_88000, 0, MOVSTR_SI_LIMIT_88000,
   0, 0, 0, MOVSTR_DI_LIMIT_88000}
};

static void block_move_loop PARAMS ((rtx, rtx, rtx, rtx, int, int));
static void block_move_no_loop PARAMS ((rtx, rtx, rtx, rtx, int, int));
static void block_move_sequence PARAMS ((rtx, rtx, rtx, rtx, int, int, int));
static void output_short_branch_defs PARAMS ((FILE *));
static int output_option PARAMS ((FILE *, const char *, const char *,
				  const char *, const char *, int, int));

/* Emit code to perform a block move.  Choose the best method.

   OPERANDS[0] is the destination.
   OPERANDS[1] is the source.
   OPERANDS[2] is the size.
   OPERANDS[3] is the alignment safe to use.  */

void
expand_block_move (dest_mem, src_mem, operands)
     rtx dest_mem;
     rtx src_mem;
     rtx *operands;
{
  int align = INTVAL (operands[3]);
  int constp = (GET_CODE (operands[2]) == CONST_INT);
  int bytes = (constp ? INTVAL (operands[2]) : 0);
  int target = (int) m88k_cpu;

  if (! (PROCESSOR_M88100 == 0
	 && PROCESSOR_M88110 == 1
	 && PROCESSOR_M88000 == 2))
    abort ();

  if (constp && bytes <= 0)
    return;

  /* Determine machine mode to do move with.  */
  if (align > 4 && !TARGET_88110)
    align = 4;
  else if (align <= 0 || align == 3)
    abort ();	/* block move invalid alignment.  */

  if (constp && bytes <= 3 * align)
    block_move_sequence (operands[0], dest_mem, operands[1], src_mem,
			 bytes, align, 0);

  else if (constp && bytes <= best_from_align[target][align])
    block_move_no_loop (operands[0], dest_mem, operands[1], src_mem,
			bytes, align);

  else if (constp && align == 4 && TARGET_88100)
    block_move_loop (operands[0], dest_mem, operands[1], src_mem,
		     bytes, align);

  else
    {
#ifdef TARGET_MEM_FUNCTIONS
      emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "memcpy"), 0,
			 VOIDmode, 3,
			 operands[0], Pmode,
			 operands[1], Pmode,
			 convert_to_mode (TYPE_MODE (sizetype), operands[2],
					  TREE_UNSIGNED (sizetype)),
			 TYPE_MODE (sizetype));
#else
      emit_library_call (gen_rtx_SYMBOL_REF (Pmode, "bcopy"), 0,
			 VOIDmode, 3,
			 operands[1], Pmode,
			 operands[0], Pmode,
			 convert_to_mode (TYPE_MODE (integer_type_node),
					  operands[2],
					  TREE_UNSIGNED (integer_type_node)),
			 TYPE_MODE (integer_type_node));
#endif
    }
}

/* Emit code to perform a block move by calling a looping movstr library
   function.  SIZE and ALIGN are known constants.  DEST and SRC are
   registers.  */

static void
block_move_loop (dest, dest_mem, src, src_mem, size, align)
     rtx dest, dest_mem;
     rtx src, src_mem;
     int size;
     int align;
{
  enum machine_mode mode;
  int count;
  int units;
  int remainder;
  rtx offset_rtx;
  rtx value_rtx;
  char entry[30];
  tree entry_name;

  /* Determine machine mode to do move with.  */
  if (align != 4)
    abort ();

  /* Determine the structure of the loop.  */
  count = size / MOVSTR_LOOP;
  units = (size - count * MOVSTR_LOOP) / align;

  if (units < 2)
    {
      count--;
      units += MOVSTR_LOOP / align;
    }

  if (count <= 0)
    {
      block_move_no_loop (dest, dest_mem, src, src_mem, size, align);
      return;
    }

  remainder = size - count * MOVSTR_LOOP - units * align;

  mode = mode_from_align[align];
  sprintf (entry, "__movstr%s%dn%d",
	   GET_MODE_NAME (mode), MOVSTR_LOOP, units * align);
  entry_name = get_identifier (entry);

  offset_rtx = GEN_INT (MOVSTR_LOOP + (1 - units) * align);

  value_rtx = gen_rtx_MEM (MEM_IN_STRUCT_P (src_mem) ? mode : BLKmode,
			   gen_rtx_PLUS (Pmode,
					 gen_rtx_REG (Pmode, 3),
					 offset_rtx));
  MEM_COPY_ATTRIBUTES (value_rtx, src_mem);

  emit_insn (gen_call_movstrsi_loop
	     (gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (entry_name)),
	      dest, src, offset_rtx, value_rtx,
	      gen_rtx_REG (mode, ((units & 1) ? 4 : 5)),
	      GEN_INT (count)));

  if (remainder)
    block_move_sequence (gen_rtx_REG (Pmode, 2), dest_mem,
			 gen_rtx_REG (Pmode, 3), src_mem,
			 remainder, align, MOVSTR_LOOP + align);
}

/* Emit code to perform a block move by calling a non-looping library
   function.  SIZE and ALIGN are known constants.  DEST and SRC are
   registers.  OFFSET is the known starting point for the output pattern.  */

static void
block_move_no_loop (dest, dest_mem, src, src_mem, size, align)
     rtx dest, dest_mem;
     rtx src, src_mem;
     int size;
     int align;
{
  enum machine_mode mode = mode_from_align[align];
  int units = size / align;
  int remainder = size - units * align;
  int most;
  int value_reg;
  rtx offset_rtx;
  rtx value_rtx;
  char entry[30];
  tree entry_name;

  if (remainder && size <= all_from_align[align])
    {
      most = all_from_align[align] - (align - remainder);
      remainder = 0;
    }
  else
    {
      most = max_from_align[align];
    }

  sprintf (entry, "__movstr%s%dx%d",
	   GET_MODE_NAME (mode), most, size - remainder);
  entry_name = get_identifier (entry);

  offset_rtx = GEN_INT (most - (size - remainder));

  value_rtx = gen_rtx_MEM (MEM_IN_STRUCT_P (src_mem) ? mode : BLKmode,
			   gen_rtx_PLUS (Pmode,
					 gen_rtx_REG (Pmode, 3),
					 offset_rtx));

  MEM_COPY_ATTRIBUTES (value_rtx, src_mem);

  value_reg = ((((most - (size - remainder)) / align) & 1) == 0
	       ? (align == 8 ? 6 : 5) : 4);

  emit_insn (gen_call_block_move
	     (gen_rtx_SYMBOL_REF (Pmode, IDENTIFIER_POINTER (entry_name)),
	      dest, src, offset_rtx, value_rtx,
	      gen_rtx_REG (mode, value_reg)));

  if (remainder)
    block_move_sequence (gen_rtx_REG (Pmode, 2), dest_mem,
			 gen_rtx_REG (Pmode, 3), src_mem,
			 remainder, align, most);
}

/* Emit code to perform a block move with an offset sequence of ld/st
   instructions (..., ld 0, st 1, ld 1, st 0, ...).  SIZE and ALIGN are
   known constants.  DEST and SRC are registers.  OFFSET is the known
   starting point for the output pattern.  */

static void
block_move_sequence (dest, dest_mem, src, src_mem, size, align, offset)
     rtx dest, dest_mem;
     rtx src, src_mem;
     int size;
     int align;
     int offset;
{
  rtx temp[2];
  enum machine_mode mode[2];
  int amount[2];
  int active[2];
  int phase = 0;
  int next;
  int offset_ld = offset;
  int offset_st = offset;

  active[0] = active[1] = FALSE;

  /* Establish parameters for the first load and for the second load if
     it is known to be the same mode as the first.  */
  amount[0] = amount[1] = align;
  mode[0] = mode_from_align[align];
  temp[0] = gen_reg_rtx (mode[0]);
  if (size >= 2 * align)
    {
      mode[1] = mode[0];
      temp[1] = gen_reg_rtx (mode[1]);
    }

  do
    {
      rtx srcp, dstp;
      next = phase;
      phase = !phase;

      if (size > 0)
	{
	  /* Change modes as the sequence tails off.  */
	  if (size < amount[next])
	    {
	      amount[next] = (size >= 4 ? 4 : (size >= 2 ? 2 : 1));
	      mode[next] = mode_from_align[amount[next]];
	      temp[next] = gen_reg_rtx (mode[next]);
	    }
	  size -= amount[next];
	  srcp = gen_rtx_MEM (MEM_IN_STRUCT_P (src_mem) ? mode[next] : BLKmode,
			      plus_constant (src, offset_ld));

	  MEM_COPY_ATTRIBUTES (srcp, src_mem);
	  emit_insn (gen_rtx_SET (VOIDmode, temp[next], srcp));
	  offset_ld += amount[next];
	  active[next] = TRUE;
	}

      if (active[phase])
	{
	  active[phase] = FALSE;
	  dstp
	    = gen_rtx_MEM (MEM_IN_STRUCT_P (dest_mem) ? mode[phase] : BLKmode,
			   plus_constant (dest, offset_st));

	  MEM_COPY_ATTRIBUTES (dstp, dest_mem);
	  emit_insn (gen_rtx_SET (VOIDmode, dstp, temp[phase]));
	  offset_st += amount[phase];
	}
    }
  while (active[next]);
}

/* Emit the code to do an AND operation.  */

const char *
output_and (operands)
     rtx operands[];
{
  unsigned int value;

  if (REG_P (operands[2]))
    return "and %0,%1,%2";

  value = INTVAL (operands[2]);
  if (SMALL_INTVAL (value))
    return "mask %0,%1,%2";
  else if ((value & 0xffff0000) == 0xffff0000)
    return "and %0,%1,%x2";
  else if ((value & 0xffff) == 0xffff)
    return "and.u %0,%1,%X2";
  else if ((value & 0xffff) == 0)
    return "mask.u %0,%1,%X2";
  else if (integer_ok_for_set (~value))
    return "clr %0,%1,%S2";
  else
    return "and.u %0,%1,%X2\n\tand %0,%0,%x2";
}

/* Emit the code to do an inclusive OR operation.  */

const char *
output_ior (operands)
     rtx operands[];
{
  unsigned int value;

  if (REG_P (operands[2]))
    return "or %0,%1,%2";

  value = INTVAL (operands[2]);
  if (SMALL_INTVAL (value))
    return "or %0,%1,%2";
  else if ((value & 0xffff) == 0)
    return "or.u %0,%1,%X2";
  else if (integer_ok_for_set (value))
    return "set %0,%1,%s2";
  else
    return "or.u %0,%1,%X2\n\tor %0,%0,%x2";
}

/* Emit the instructions for doing an XOR.  */

const char *
output_xor (operands)
     rtx operands[];
{
  unsigned int value;

  if (REG_P (operands[2]))
    return "xor %0,%1,%2";

  value = INTVAL (operands[2]);
  if (SMALL_INTVAL (value))
    return "xor %0,%1,%2";
  else if ((value & 0xffff) == 0)
    return "xor.u %0,%1,%X2";
  else
    return "xor.u %0,%1,%X2\n\txor %0,%0,%x2";
}

/* Output a call.  Normally this is just bsr or jsr, but this also deals with
   accomplishing a branch after the call by incrementing r1.  This requires
   that various assembler bugs be accommodated.  The 4.30 DG/UX assembler
   requires that forward references not occur when computing the difference of
   two labels.  The [version?] Motorola assembler computes a word difference.
   No doubt there's more to come!

   It would seem the same idea could be used to tail call, but in this case,
   the epilogue will be non-null.  */

static rtx sb_name = 0;
static rtx sb_high = 0;
static rtx sb_low = 0;

const char *
output_call (operands, addr)
     rtx operands[];
     rtx addr;
{
  operands[0] = addr;
  if (final_sequence)
    {
      rtx jump;
      rtx seq_insn;

      /* This can be generalized, but there is currently no need.  */
      if (XVECLEN (final_sequence, 0) != 2)
	abort ();

      /* The address of interior insns is not computed, so use the sequence.  */
      seq_insn = NEXT_INSN (PREV_INSN (XVECEXP (final_sequence, 0, 0)));
      jump = XVECEXP (final_sequence, 0, 1);
      if (GET_CODE (jump) == JUMP_INSN)
	{
	  rtx low, high;
	  const char *last;
	  rtx dest = XEXP (SET_SRC (PATTERN (jump)), 0);
	  int delta = 4 * (INSN_ADDRESSES (INSN_UID (dest))
			   - INSN_ADDRESSES (INSN_UID (seq_insn))
			   - 2);
#if (MONITOR_GCC & 0x2) /* How often do long branches happen?  */
	  if ((unsigned) (delta + 0x8000) >= 0x10000)
	    warning ("internal gcc monitor: short-branch(%x)", delta);
#endif

	  /* Delete the jump.  */
	  PUT_CODE (jump, NOTE);
	  NOTE_LINE_NUMBER (jump) = NOTE_INSN_DELETED;
	  NOTE_SOURCE_FILE (jump) = 0;

	  /* We only do this optimization if -O2, modifying the value of
	     r1 in the delay slot confuses debuggers and profilers on some
	     systems.

	     If we loose, we must use the non-delay form.  This is unlikely
	     to ever happen.  If it becomes a problem, claim that a call
	     has two delay slots and only the second can be filled with
	     a jump.  

	     The 88110 can lose when a jsr.n r1 is issued and a page fault
	     occurs accessing the delay slot.  So don't use jsr.n form when
	     jumping thru r1.
	   */
#ifdef AS_BUG_IMMEDIATE_LABEL /* The assembler restricts immediate values.  */
	  if (optimize < 2
	      || ! ADD_INTVAL (delta * 2)
#else
	  if (optimize < 2
	      || ! ADD_INTVAL (delta)
#endif