pa.c

早期freebsd实现

      if (GET_CODE (operands[0]) == REG)
	{
	  rtx xoperands[3];
	  xoperands[2] = operands[1];
	  xoperands[1] = gen_rtx (REG, SImode, REGNO (operands[0]) + 1);
	  xoperands[0] = operands[0];
	  output_asm_insn
	    ("fstds %2,-16(0,30)\n\tldw -12(0,30),%1\n\tldw -16(0,30),%0",
	     xoperands);
	}
      else
	output_asm_insn ("fstds%F0 %1,%0", operands);
    }
  else abort ();
  return "";
}

/* Return a REG that occurs in ADDR with coefficient 1.
   ADDR can be effectively incremented by incrementing REG.  */

static rtx
find_addr_reg (addr)
     rtx addr;
{
  while (GET_CODE (addr) == PLUS)
    {
      if (GET_CODE (XEXP (addr, 0)) == REG)
	addr = XEXP (addr, 0);
      else if (GET_CODE (XEXP (addr, 1)) == REG)
	addr = XEXP (addr, 1);
      else if (CONSTANT_P (XEXP (addr, 0)))
	addr = XEXP (addr, 1);
      else if (CONSTANT_P (XEXP (addr, 1)))
	addr = XEXP (addr, 0);
      else
	abort ();
    }
  if (GET_CODE (addr) == REG)
    return addr;
  abort ();
}

/* Emit code to perform a block move.

   Restriction: If the length argument is non-constant, alignment
   must be 4.

   OPERANDS[0] is the destination pointer as a REG, clobbered.
   OPERANDS[1] is the source pointer as a REG, clobbered.
   if SIZE_IS_CONSTANT
     OPERANDS[2] is a register for temporary storage.
     OPERANDS[4] is the size as a CONST_INT
   else
     OPERANDS[2] is a REG which will contain the size, clobbered.
   OPERANDS[3] is a register for temporary storage.
   OPERANDS[5] is the alignment safe to use, as a CONST_INT.  */

char *
output_block_move (operands, size_is_constant)
     rtx *operands;
     int size_is_constant;
{
  int align = INTVAL (operands[5]);
  unsigned long n_bytes;

  /* We can't move more than four bytes at a time because the PA
     has no longer integer move insns.  (Could use fp mem ops?)  */
  if (align > 4)
    align = 4;

  if (size_is_constant)
    {
      unsigned long n_items;
      unsigned long offset;
      rtx temp;

      n_bytes = INTVAL (operands[4]);
      if (n_bytes == 0)
	return "";

      if (align >= 4)
	{
	  /* Don't unroll too large blocks.  */
	  if (n_bytes > 64)
	    goto copy_with_loop;

	  /* Read and store using two registers, and hide latency
	     by deferring the stores until three instructions after
	     the corresponding load.  The last load insn will read
	     the entire word were the last bytes are, possibly past
	     the end of the source block, but since loads are aligned,
	     this is harmless.  */

	  output_asm_insn ("ldws,ma 4(0,%1),%2", operands);

	  for (offset = 4; offset < n_bytes; offset += 4)
	    {
	      output_asm_insn ("ldws,ma 4(0,%1),%3", operands);
	      output_asm_insn ("stws,ma %2,4(0,%0)", operands);

	      temp = operands[2];
	      operands[2] = operands[3];
	      operands[3] = temp;
	    }
	  if (n_bytes % 4 == 0)
	    /* Store the last word.  */
	    output_asm_insn ("stw %2,0(0,%0)", operands);
	  else
	    {
	      /* Store the last, partial word.  */
	      operands[4] = gen_rtx (CONST_INT, VOIDmode, n_bytes % 4);
	      output_asm_insn ("stbys,e %2,%4(0,%0)", operands);
	    }
	  return "";
	}

      if (align >= 2 && n_bytes >= 2)
	{
	  output_asm_insn ("ldhs,ma 2(0,%1),%2", operands);

	  for (offset = 2; offset + 2 <= n_bytes; offset += 2)
	    {
	      output_asm_insn ("ldhs,ma 2(0,%1),%3", operands);
	      output_asm_insn ("sths,ma %2,2(0,%0)", operands);

	      temp = operands[2];
	      operands[2] = operands[3];
	      operands[3] = temp;
	    }
	  if (n_bytes % 2 != 0)
	    output_asm_insn ("ldb 0(0,%1),%3", operands);

	  output_asm_insn ("sths,ma %2,2(0,%0)", operands);

	  if (n_bytes % 2 != 0)
	    output_asm_insn ("stb %3,0(0,%0)", operands);

	  return "";
	}

      output_asm_insn ("ldbs,ma 1(0,%1),%2", operands);

      for (offset = 1; offset + 1 <= n_bytes; offset += 1)
	{
	  output_asm_insn ("ldbs,ma 1(0,%1),%3", operands);
	  output_asm_insn ("stbs,ma %2,1(0,%0)", operands);

	  temp = operands[2];
	  operands[2] = operands[3];
	  operands[3] = temp;
	}
      output_asm_insn ("stb %2,0(0,%0)", operands);

      return "";
    }

  if (align != 4)
    abort();
     
 copy_with_loop:

  if (size_is_constant)
    {
      /* Size is compile-time determined, and also not
	 very small (such small cases are handled above).  */
      operands[4] = gen_rtx (CONST_INT, VOIDmode, n_bytes - 4);
      output_asm_insn ("ldo %4(0),%2", operands);
    }
  else
    {
      /* Decrement counter by 4, and if it becomes negative, jump past the
	 word copying loop.  */
      output_asm_insn ("addib,<,n -4,%2,.+16", operands);
    }

  /* Copying loop.  Note that the first load is in the annulled delay slot
     of addib.  Is it OK on PA to have a load in a delay slot, i.e. is a
     possible page fault stopped in time?  */
  output_asm_insn ("ldws,ma 4(0,%1),%3", operands);
  output_asm_insn ("addib,>= -4,%2,.-4", operands);
  output_asm_insn ("stws,ma %3,4(0,%0)", operands);

  /* The counter is negative, >= -4.  The remaining number of bytes are
     determined by the two least significant bits.  */

  if (size_is_constant)
    {
      if (n_bytes % 4 != 0)
	{
	  /* Read the entire word of the source block tail.  */
	  output_asm_insn ("ldw 0(0,%1),%3", operands);
	  operands[4] = gen_rtx (CONST_INT, VOIDmode, n_bytes % 4);
	  output_asm_insn ("stbys,e %3,%4(0,%0)", operands);
	}
    }
  else
    {
      /* Add 4 to counter.  If it becomes zero, we're done.  */
      output_asm_insn ("addib,=,n 4,%2,.+16", operands);

      /* Read the entire word of the source block tail.  (Also this
	 load is in an annulled delay slot.)  */
      output_asm_insn ("ldw 0(0,%1),%3", operands);

      /* Make %0 point at the first byte after the destination block.  */
      output_asm_insn ("add %2,%0,%0", operands);
      /* Store the leftmost bytes, up to, but not including, the address
	 in %0.  */
      output_asm_insn ("stbys,e %3,0(0,%0)", operands);
    }
  return "";
}


char *
output_and (operands)
     rtx *operands;
{
  if (GET_CODE (operands[2]) == CONST_INT && INTVAL (operands[2]) != 0)
    {
      unsigned mask = INTVAL (operands[2]);
      int ls0, ls1, ms0, p, len;

      for (ls0 = 0; ls0 < 32; ls0++)
	if ((mask & (1 << ls0)) == 0)
	  break;

      for (ls1 = ls0; ls1 < 32; ls1++)
	if ((mask & (1 << ls1)) != 0)
	  break;

      for (ms0 = ls1; ms0 < 32; ms0++)
	if ((mask & (1 << ms0)) == 0)
	  break;

      if (ms0 != 32)
	abort();

      if (ls1 == 32)
	{
	  len = ls0;

	  if (len == 0)
	    abort ();

	  operands[2] = gen_rtx (CONST_INT, VOIDmode, len);
	  return "extru %1,31,%2,%0";
	}
      else
	{
	  /* We could use this `depi' for the case above as well, but `depi'
	     requires one more register file access than an `extru'.  */

	  p = 31 - ls0;
	  len = ls1 - ls0;

	  operands[2] = gen_rtx (CONST_INT, VOIDmode, p);
	  operands[3] = gen_rtx (CONST_INT, VOIDmode, len);
	  return "depi 0,%2,%3,%0";
	}
    }
  else
    return "and %1,%2,%0";
}

char *
output_ior (operands)
     rtx *operands;
{
  if (GET_CODE (operands[2]) == CONST_INT)
    {
      unsigned mask = INTVAL (operands[2]);
      int bs0, bs1, bs2, p, len;
 
      if (INTVAL (operands[2]) == 0)
	return "copy %1,%0";

      for (bs0 = 0; bs0 < 32; bs0++)
	if ((mask & (1 << bs0)) != 0)
	  break;

      for (bs1 = bs0; bs1 < 32; bs1++)
	if ((mask & (1 << bs1)) == 0)
	  break;

      if (bs1 != 32 && ((unsigned) 1 << bs1) <= mask)
	abort();

      p = 31 - bs0;
      len = bs1 - bs0;

      operands[2] = gen_rtx (CONST_INT, VOIDmode, p);
      operands[3] = gen_rtx (CONST_INT, VOIDmode, len);
      return "depi -1,%2,%3,%0";
    }
  else
    return "or %1,%2,%0";
}

/* Output an ascii string.  */
output_ascii (file, p, size)
     FILE *file;
     unsigned char *p;
     int size;
{
  int i;
  int chars_output;
  unsigned char partial_output[16];	/* Max space 4 chars can occupy.   */

  /* The HP assembler can only take strings of 256 characters at one
     time.  This is a limitation on input line length, *not* the
     length of the string.  Sigh.  Even worse, it seems that the
     restriction is in number of input characters (see \xnn &
     \whatever).  So we have to do this very carefully.  */

  fprintf (file, "\t.STRING \"");

  chars_output = 0;
  for (i = 0; i < size; i += 4)
    {
      int co = 0;
      int io = 0;
      for (io = 0, co = 0; io < MIN (4, size - i); io++)
	{
	  register unsigned int c = p[i + io];

	  if (c == '\"' || c == '\\')
	    partial_output[co++] = '\\';
	  if (c >= ' ' && c < 0177)
	    partial_output[co++] = c;
	  else
	    {
	      unsigned int hexd;
	      partial_output[co++] = '\\';
	      partial_output[co++] = 'x';
	      hexd =  c  / 16 - 0 + '0';
	      if (hexd > '9')
		hexd -= '9' - 'a' + 1;
	      partial_output[co++] = hexd;
	      hexd =  c % 16 - 0 + '0';
	      if (hexd > '9')
		hexd -= '9' - 'a' + 1;
	      partial_output[co++] = hexd;
	    }
	}
      if (chars_output + co > 243)
	{
	  fprintf (file, "\"\n\t.STRING \"");
	  chars_output = 0;
	}
      fwrite (partial_output, 1, co, file);
      chars_output += co;
      co = 0;
    }
  fprintf (file, "\"\n");
}

/* You may have trouble believing this, but this is the HP825 stack
   layout.  Wow.

   Offset		Contents

   Variable arguments	(optional; any number may be allocated)

   SP-(4*(N+9))		arg word N
   	:		    :
      SP-56		arg word 5
      SP-52		arg word 4

   Fixed arguments	(must be allocated; may remain unused)

      SP-48		arg word 3
      SP-44		arg word 2
      SP-40		arg word 1
      SP-36		arg word 0

   Frame Marker

      SP-32		External Data Pointer (DP)
      SP-28		External sr4
      SP-24		External/stub RP (RP')
      SP-20		Current RP
      SP-16		Static Link
      SP-12		Clean up
      SP-8		Calling Stub RP (RP'')
      SP-4		Previous SP

   Top of Frame

      SP-0		Stack Pointer (points to next available address)

*/

/* This function saves registers as follows.  Registers marked with ' are
   this function's registers (as opposed to the previous function's).
   If a frame_pointer isn't needed, r4 is saved as a general register;
   the space for the frame pointer is still allocated, though, to keep
   things simple.


   Top of Frame

       SP (FP')		Previous FP
       SP + 4		Alignment filler (sigh)
       SP + 8		Space for locals reserved here.
       .
       .
       .
       SP + n		All call saved register used.
       .
       .
       .
       SP + o		All call saved fp registers used.
       .
       .
       .
       SP + p (SP')	points to next available address.
       
*/

/* Helper functions */
void
print_stw (file, r, disp, base)
     FILE *file;
     int r, disp, base;
{
  if (VAL_14_BITS_P (disp))
    fprintf (file, "\tstw %d,%d(0,%d)\n", r, disp, base);
  else
    fprintf (file, "\taddil L'%d,%d\n\tstw %d,R'%d(0,1)\n", disp, base,
	     r, disp);
}

void
print_ldw (file, r, disp, base)
     FILE *file;
     int r, disp, base;
{
  if (VAL_14_BITS_P (disp))
    fprintf (file, "\tldw %d(0,%d),%d\n", disp, base, r);
  else
    fprintf (file, "\taddil L'%d,%d\n\tldw R'%d(0,1),%d\n", disp, base,
	     disp, r);
}

/* Global variables set by FUNCTION_PROLOGUE.  */
/* Size of frame.  Need to know this to emit return insns from
   leaf procedures.  */
int apparent_fsize;
int actual_fsize;
int local_fsize, save_fregs;

int
compute_frame_size (size, leaf_function, fregs_live)
     int size;
     int leaf_function;
     int *fregs_live;
{
  extern int current_function_outgoing_args_size;
  int i;

  /* 8 is space for frame pointer + filler */
  local_fsize = actual_fsize = size + 8;

  /* fp is stored in a special place. */
  for (i = 18; i >= 5; i--)
    if (regs_ever_live[i])
      actual_fsize += 4;

  if (regs_ever_live[3])
    actual_fsize += 4;
  actual_fsize = (actual_fsize + 7) & ~7;

  if (!TARGET_SNAKE)
    {
      for (i = 47; i >= 44; i--)
	if (regs_ever_live[i])
	  {
	    actual_fsize += 8;
	    if (fregs_live)
	      *fregs_live = 1;
	  }
    }
  else
    {
      for (i = 90; i >= 72; i -= 2)
	if (regs_ever_live[i] || regs_ever_live[i + 1])
	  {
	    actual_fsize += 8;
	    if (fregs_live)
	      *fregs_live = 1;
	  }
    }
  return actual_fsize + current_function_outgoing_args_size;
}
     
void
output_function_prologue (file, size, leaf_function)
     FILE *file;
     int size;
     int leaf_function;
{
  extern char call_used_regs[];
  extern int frame_pointer_needed;
  extern int current_function_returns_struct;
  int i, offset;

  save_fregs = 0;