alpha.c

gcc-2.95.3 Linux下最常用的C编译器

          else if (a >= 4 && c % 4 == 0)
	    dst_align = 4;
          else if (a >= 2 && c % 2 == 0)
	    dst_align = 2;
	}
    }

  /*
   * Load the entire block into registers.
   */

  if (GET_CODE (XEXP (orig_src, 0)) == ADDRESSOF)
    {
      enum machine_mode mode;
      tmp = XEXP (XEXP (orig_src, 0), 0);

      /* Don't use the existing register if we're reading more than
	 is held in the register.  Nor if there is not a mode that
	 handles the exact size.  */
      mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 1);
      if (mode != BLKmode
	  && GET_MODE_SIZE (GET_MODE (tmp)) >= bytes)
	{
	  if (mode == TImode)
	    {
	      data_regs[nregs] = gen_lowpart (DImode, tmp);
	      data_regs[nregs+1] = gen_highpart (DImode, tmp);
	      nregs += 2;
	    }
	  else
	    data_regs[nregs++] = gen_lowpart (mode, tmp);
	  goto src_done;
	}

      /* No appropriate mode; fall back on memory.  */
      orig_src = change_address (orig_src, GET_MODE (orig_src),
				 copy_addr_to_reg (XEXP (orig_src, 0)));
    }

  ofs = 0;
  if (src_align >= 8 && bytes >= 8)
    {
      words = bytes / 8;

      for (i = 0; i < words; ++i)
	data_regs[nregs+i] = gen_reg_rtx(DImode);

      for (i = 0; i < words; ++i)
	{
	  emit_move_insn (data_regs[nregs+i],
			  change_address (orig_src, DImode,
					  plus_constant (XEXP (orig_src, 0),
							 ofs + i*8)));
	}

      nregs += words;
      bytes -= words * 8;
      ofs += words * 8;
    }
  if (src_align >= 4 && bytes >= 4)
    {
      words = bytes / 4;

      for (i = 0; i < words; ++i)
	data_regs[nregs+i] = gen_reg_rtx(SImode);

      for (i = 0; i < words; ++i)
	{
	  emit_move_insn (data_regs[nregs+i],
			  change_address (orig_src, SImode,
					  plus_constant (XEXP (orig_src, 0),
							 ofs + i*4)));
	}

      nregs += words;
      bytes -= words * 4;
      ofs += words * 4;
    }
  if (bytes >= 16)
    {
      words = bytes / 8;

      for (i = 0; i < words+1; ++i)
	data_regs[nregs+i] = gen_reg_rtx(DImode);

      alpha_expand_unaligned_load_words (data_regs + nregs, orig_src,
					 words, ofs);

      nregs += words;
      bytes -= words * 8;
      ofs += words * 8;
    }
  if (!TARGET_BWX && bytes >= 8)
    {
      data_regs[nregs++] = tmp = gen_reg_rtx (DImode);
      alpha_expand_unaligned_load (tmp, orig_src, 8, ofs, 0);
      bytes -= 8;
      ofs += 8;
    }
  if (!TARGET_BWX && bytes >= 4)
    {
      data_regs[nregs++] = tmp = gen_reg_rtx (SImode);
      alpha_expand_unaligned_load (tmp, orig_src, 4, ofs, 0);
      bytes -= 4;
      ofs += 4;
    }
  if (bytes >= 2)
    {
      if (src_align >= 2)
	{
	  do {
	    data_regs[nregs++] = tmp = gen_reg_rtx (HImode);
	    emit_move_insn (tmp,
			    change_address (orig_src, HImode,
					    plus_constant (XEXP (orig_src, 0),
							   ofs)));
	    bytes -= 2;
	    ofs += 2;
	  } while (bytes >= 2);
	}
      else if (!TARGET_BWX)
	{
	  data_regs[nregs++] = tmp = gen_reg_rtx (HImode);
	  alpha_expand_unaligned_load (tmp, orig_src, 2, ofs, 0);
	  bytes -= 2;
	  ofs += 2;
	}
    }
  while (bytes > 0)
    {
      data_regs[nregs++] = tmp = gen_reg_rtx (QImode);
      emit_move_insn (tmp,
		      change_address (orig_src, QImode,
				      plus_constant (XEXP (orig_src, 0),
						     ofs)));
      bytes -= 1;
      ofs += 1;
    }
 src_done:

  if (nregs > (int)(sizeof(data_regs)/sizeof(*data_regs)))
    abort();

  /*
   * Now save it back out again.
   */

  i = 0, ofs = 0;

  if (GET_CODE (XEXP (orig_dst, 0)) == ADDRESSOF)
    {
      enum machine_mode mode;
      tmp = XEXP (XEXP (orig_dst, 0), 0);

      mode = mode_for_size (orig_bytes * BITS_PER_UNIT, MODE_INT, 1);
      if (GET_MODE (tmp) == mode)
	{
	  if (nregs == 1)
	    {
	      emit_move_insn (tmp, data_regs[0]);
	      i = 1;
	      goto dst_done;
	    }
	  else if (nregs == 2 && mode == TImode)
	    {
	      /* Undo the subregging done above when copying between
		 two TImode registers.  */
	      if (GET_CODE (data_regs[0]) == SUBREG
		  && GET_MODE (SUBREG_REG (data_regs[0])) == TImode)
		{
		  emit_move_insn (tmp, SUBREG_REG (data_regs[0]));
		}
	      else
		{
		  rtx seq;

		  start_sequence ();
		  emit_move_insn (gen_lowpart (DImode, tmp), data_regs[0]);
		  emit_move_insn (gen_highpart (DImode, tmp), data_regs[1]);
		  seq = get_insns ();
		  end_sequence ();

		  emit_no_conflict_block (seq, tmp, data_regs[0],
					  data_regs[1], NULL_RTX);
		}

	      i = 2;
	      goto dst_done;
	    }
	}

      /* ??? If nregs > 1, consider reconstructing the word in regs.  */
      /* ??? Optimize mode < dst_mode with strict_low_part.  */

      /* No appropriate mode; fall back on memory.  We can speed things
	 up by recognizing extra alignment information.  */
      orig_dst = change_address (orig_dst, GET_MODE (orig_dst),
				 copy_addr_to_reg (XEXP (orig_dst, 0)));
      dst_align = GET_MODE_SIZE (GET_MODE (tmp));
    }

  /* Write out the data in whatever chunks reading the source allowed.  */
  if (dst_align >= 8)
    {
      while (i < nregs && GET_MODE (data_regs[i]) == DImode)
	{
	  emit_move_insn (change_address (orig_dst, DImode,
					  plus_constant (XEXP (orig_dst, 0),
							 ofs)),
			  data_regs[i]);
	  ofs += 8;
	  i++;
	}
    }
  if (dst_align >= 4)
    {
      /* If the source has remaining DImode regs, write them out in
	 two pieces.  */
      while (i < nregs && GET_MODE (data_regs[i]) == DImode)
	{
	  tmp = expand_binop (DImode, lshr_optab, data_regs[i], GEN_INT (32),
			      NULL_RTX, 1, OPTAB_WIDEN);

	  emit_move_insn (change_address (orig_dst, SImode,
					  plus_constant (XEXP (orig_dst, 0),
							 ofs)),
			  gen_lowpart (SImode, data_regs[i]));
	  emit_move_insn (change_address (orig_dst, SImode,
					  plus_constant (XEXP (orig_dst, 0),
							 ofs+4)),
			  gen_lowpart (SImode, tmp));
	  ofs += 8;
	  i++;
	}

      while (i < nregs && GET_MODE (data_regs[i]) == SImode)
	{
	  emit_move_insn (change_address(orig_dst, SImode,
					 plus_constant (XEXP (orig_dst, 0),
							ofs)),
			  data_regs[i]);
	  ofs += 4;
	  i++;
	}
    }
  if (i < nregs && GET_MODE (data_regs[i]) == DImode)
    {
      /* Write out a remaining block of words using unaligned methods.  */

      for (words = 1; i+words < nregs ; ++words)
	if (GET_MODE (data_regs[i+words]) != DImode)
	  break;

      if (words == 1)
	alpha_expand_unaligned_store (orig_dst, data_regs[i], 8, ofs);
      else
        alpha_expand_unaligned_store_words (data_regs+i, orig_dst, words, ofs);
     
      i += words;
      ofs += words * 8;
    }

  /* Due to the above, this won't be aligned.  */
  /* ??? If we have more than one of these, consider constructing full
     words in registers and using alpha_expand_unaligned_store_words.  */
  while (i < nregs && GET_MODE (data_regs[i]) == SImode)
    {
      alpha_expand_unaligned_store (orig_dst, data_regs[i], 4, ofs);
      ofs += 4;
      i++;
    }

  if (dst_align >= 2)
    while (i < nregs && GET_MODE (data_regs[i]) == HImode)
      {
	emit_move_insn (change_address (orig_dst, HImode,
					plus_constant (XEXP (orig_dst, 0),
						       ofs)),
			data_regs[i]);
	i++;
	ofs += 2;
      }
  else
    while (i < nregs && GET_MODE (data_regs[i]) == HImode)
      {
	alpha_expand_unaligned_store (orig_dst, data_regs[i], 2, ofs);
	i++;
	ofs += 2;
      }
  while (i < nregs && GET_MODE (data_regs[i]) == QImode)
    {
      emit_move_insn (change_address (orig_dst, QImode,
				      plus_constant (XEXP (orig_dst, 0),
						     ofs)),
		      data_regs[i]);
      i++;
      ofs += 1;
    }
 dst_done:

  if (i != nregs)
    abort();

  return 1;
}

int
alpha_expand_block_clear (operands)
     rtx operands[];
{
  rtx bytes_rtx	= operands[1];
  rtx align_rtx = operands[2];
  HOST_WIDE_INT bytes = INTVAL (bytes_rtx);
  HOST_WIDE_INT align = INTVAL (align_rtx);
  rtx orig_dst	= operands[0];
  rtx tmp;
  HOST_WIDE_INT i, words, ofs = 0;
  
  if (bytes <= 0)
    return 1;
  if (bytes > MAX_MOVE_WORDS*8)
    return 0;

  /* Look for stricter alignment.  */

  tmp = XEXP (orig_dst, 0);
  if (GET_CODE (tmp) == REG)
    {
      if (REGNO_POINTER_ALIGN (REGNO (tmp)) > align)
	align = REGNO_POINTER_ALIGN (REGNO (tmp));
    }
  else if (GET_CODE (tmp) == PLUS
	   && GET_CODE (XEXP (tmp, 0)) == REG
	   && GET_CODE (XEXP (tmp, 1)) == CONST_INT)
    {
      HOST_WIDE_INT c = INTVAL (XEXP (tmp, 1));
      int a = REGNO_POINTER_ALIGN (REGNO (XEXP (tmp, 0)));

      if (a > align)
	{
          if (a >= 8 && c % 8 == 0)
	    align = 8;
          else if (a >= 4 && c % 4 == 0)
	    align = 4;
          else if (a >= 2 && c % 2 == 0)
	    align = 2;
	}
    }
  else if (GET_CODE (tmp) == ADDRESSOF)
    {
      enum machine_mode mode;

      mode = mode_for_size (bytes * BITS_PER_UNIT, MODE_INT, 1);
      if (GET_MODE (XEXP (tmp, 0)) == mode)
	{
	  emit_move_insn (XEXP (tmp, 0), const0_rtx);
	  return 1;
	}

      /* No appropriate mode; fall back on memory.  */
      orig_dst = change_address (orig_dst, GET_MODE (orig_dst),
				 copy_addr_to_reg (tmp));
      align = GET_MODE_SIZE (GET_MODE (XEXP (tmp, 0)));
    }

  /* Handle a block of contiguous words first.  */

  if (align >= 8 && bytes >= 8)
    {
      words = bytes / 8;

      for (i = 0; i < words; ++i)
	{
	  emit_move_insn (change_address(orig_dst, DImode,
					 plus_constant (XEXP (orig_dst, 0),
							ofs + i*8)),
			  const0_rtx);
	}

      bytes -= words * 8;
      ofs += words * 8;
    }
  if (align >= 4 && bytes >= 4)
    {
      words = bytes / 4;

      for (i = 0; i < words; ++i)
	{
	  emit_move_insn (change_address (orig_dst, SImode,
					  plus_constant (XEXP (orig_dst, 0),
							 ofs + i*4)),
			  const0_rtx);
	}

      bytes -= words * 4;
      ofs += words * 4;
    }
  if (bytes >= 16)
    {
      words = bytes / 8;

      alpha_expand_unaligned_store_words (NULL, orig_dst, words, ofs);

      bytes -= words * 8;
      ofs += words * 8;
    }

  /* Next clean up any trailing pieces.  We know from the contiguous
     block move that there are no aligned SImode or DImode hunks left.  */

  if (!TARGET_BWX && bytes >= 8)
    {
      alpha_expand_unaligned_store (orig_dst, const0_rtx, 8, ofs);
      bytes -= 8;
      ofs += 8;
    }
  if (!TARGET_BWX && bytes >= 4)
    {
      alpha_expand_unaligned_store (orig_dst, const0_rtx, 4, ofs);
      bytes -= 4;
      ofs += 4;
    }
  if (bytes >= 2)
    {
      if (align >= 2)
	{
	  do {
	    emit_move_insn (change_address (orig_dst, HImode,
					    plus_constant (XEXP (orig_dst, 0),
							   ofs)),
			    const0_rtx);
	    bytes -= 2;
	    ofs += 2;
	  } while (bytes >= 2);
	}
      else if (!TARGET_BWX)
	{
	  alpha_expand_unaligned_store (orig_dst, const0_rtx, 2, ofs);
	  bytes -= 2;
	  ofs += 2;
	}
    }
  while (bytes > 0)
    {
      emit_move_insn (change_address (orig_dst, QImode,
				      plus_constant (XEXP (orig_dst, 0),
						     ofs)),
		      const0_rtx);
      bytes -= 1;
      ofs += 1;
    }

  return 1;
}


/* Adjust the cost of a scheduling dependency.  Return the new cost of
   a dependency LINK or INSN on DEP_INSN.  COST is the current cost.  */

int
alpha_adjust_cost (insn, link, dep_insn, cost)
     rtx insn;
     rtx link;
     rtx dep_insn;
     int cost;
{
  rtx set, set_src;
  enum attr_type insn_type, dep_insn_type;

  /* If the dependence is an anti-dependence, there is no cost.  For an
     output dependence, there is sometimes a cost, but it doesn't seem
     worth handling those few cases.  */

  if (REG_NOTE_KIND (link) != 0)
    return 0;

  /* If we can't recognize the insns, we can't really do anything.  */
  if (recog_memoized (insn) < 0 || recog_memoized (dep_insn) < 0)
    return cost;

  insn_type = get_attr_type (insn);
  dep_insn_type = get_attr_type (dep_insn);

  /* Bring in the user-defined memory latency.  */
  if (dep_insn_type == TYPE_ILD
      || dep_insn_type == TYPE_FLD