sh.c

GUN开源阻止下的编译器GCC

   Worst case code looks like:

   mov.l L1,rn
   bra   L2
   nop
   align
   L1:   .long value
   L2:
   ..

   mov.l L3,rn
   bra   L4
   nop
   align
   L3:   .long value
   L4:
   ..

   We fix this by performing a scan before scheduling, which notices which
   instructions need to have their operands fetched from the constant table
   and builds the table.

   The algorithm is:

   scan, find an instruction which needs a pcrel move.  Look forward, find the
   last barrier which is within MAX_COUNT bytes of the requirement.
   If there isn't one, make one.  Process all the instructions between
   the find and the barrier.

   In the above example, we can tell that L3 is within 1k of L1, so
   the first move can be shrunk from the 3 insn+constant sequence into
   just 1 insn, and the constant moved to L3 to make:

   mov.l        L1,rn
   ..
   mov.l        L3,rn
   bra          L4
   nop
   align
   L3:.long value
   L4:.long value

   Then the second move becomes the target for the shortening process.  */

typedef struct
{
  rtx value;			/* Value in table.  */
  rtx label;			/* Label of value.  */
  enum machine_mode mode;	/* Mode of value.  */
} pool_node;

/* The maximum number of constants that can fit into one pool, since
   the pc relative range is 0...1020 bytes and constants are at least 4
   bytes long.  */

#define MAX_POOL_SIZE (1020/4)
static pool_node pool_vector[MAX_POOL_SIZE];
static int pool_size;

/* ??? If we need a constant in HImode which is the truncated value of a
   constant we need in SImode, we could combine the two entries thus saving
   two bytes.  Is this common enough to be worth the effort of implementing
   it?  */

/* ??? This stuff should be done at the same time that we shorten branches.
   As it is now, we must assume that all branches are the maximum size, and
   this causes us to almost always output constant pools sooner than
   necessary.  */

/* Add a constant to the pool and return its label.  */

static rtx
add_constant (x, mode)
     rtx x;
     enum machine_mode mode;
{
  int i;
  rtx lab;

  /* First see if we've already got it.  */
  for (i = 0; i < pool_size; i++)
    {
      if (x->code == pool_vector[i].value->code
	  && mode == pool_vector[i].mode)
	{
	  if (x->code == CODE_LABEL)
	    {
	      if (XINT (x, 3) != XINT (pool_vector[i].value, 3))
		continue;
	    }
	  if (rtx_equal_p (x, pool_vector[i].value))
	    return pool_vector[i].label;
	}
    }

  /* Need a new one.  */
  pool_vector[pool_size].value = x;
  lab = gen_label_rtx ();
  pool_vector[pool_size].mode = mode;
  pool_vector[pool_size].label = lab;
  pool_size++;
  return lab;
}

/* Output the literal table.  */

static void
dump_table (scan)
     rtx scan;
{
  int i;
  int need_align = 1;

  /* Do two passes, first time dump out the HI sized constants.  */

  for (i = 0; i < pool_size; i++)
    {
      pool_node *p = &pool_vector[i];

      if (p->mode == HImode)
	{
	  if (need_align)
	    {
	      scan = emit_insn_after (gen_align_2 (), scan);
	      need_align = 0;
	    }
	  scan = emit_label_after (p->label, scan);
	  scan = emit_insn_after (gen_consttable_2 (p->value), scan);
	}
    }

  need_align = 1;

  for (i = 0; i < pool_size; i++)
    {
      pool_node *p = &pool_vector[i];

      switch (p->mode)
	{
	case HImode:
	  break;
	case SImode:
	  if (need_align)
	    {
	      need_align = 0;
	      scan = emit_label_after (gen_label_rtx (), scan);
	      scan = emit_insn_after (gen_align_4 (), scan);
	    }
	  scan = emit_label_after (p->label, scan);
	  scan = emit_insn_after (gen_consttable_4 (p->value), scan);
	  break;
	case DImode:
	  if (need_align)
	    {
	      need_align = 0;
	      scan = emit_label_after (gen_label_rtx (), scan);
	      scan = emit_insn_after (gen_align_4 (), scan);
	    }
	  scan = emit_label_after (p->label, scan);
	  scan = emit_insn_after (gen_consttable_8 (p->value), scan);
	  break;
	default:
	  abort ();
	  break;
	}
    }

  scan = emit_insn_after (gen_consttable_end (), scan);
  scan = emit_barrier_after (scan);
  pool_size = 0;
}

/* Return non-zero if constant would be an ok source for a
   mov.w instead of a mov.l.  */

static int
hi_const (src)
     rtx src;
{
  return (GET_CODE (src) == CONST_INT
	  && INTVAL (src) >= -32768
	  && INTVAL (src) <= 32767);
}

/* Non-zero if the insn is a move instruction which needs to be fixed.  */

/* ??? For a DImode/DFmode moves, we don't need to fix it if each half of the
   CONST_DOUBLE input value is CONST_OK_FOR_I.  For a SFmode move, we don't
   need to fix it if the input value is CONST_OK_FOR_I.  */

static int
broken_move (insn)
     rtx insn;
{
  if (GET_CODE (insn) == INSN
      && GET_CODE (PATTERN (insn)) == SET
      /* We can load any 8 bit value if we don't care what the high
	 order bits end up as.  */
      && GET_MODE (SET_DEST (PATTERN (insn))) != QImode
      && CONSTANT_P (SET_SRC (PATTERN (insn)))
      && (GET_CODE (SET_SRC (PATTERN (insn))) != CONST_INT
	  || ! CONST_OK_FOR_I (INTVAL (SET_SRC (PATTERN (insn))))))
    return 1;

  return 0;
}

/* Find the last barrier from insn FROM which is close enough to hold the
   constant pool.  If we can't find one, then create one near the end of
   the range.  */

/* ??? It would be good to put constant pool tables between a case jump and
   the jump table.  This fails for two reasons.  First, there is no
   barrier after the case jump.  This is a bug in the casesi pattern.
   Second, inserting the table here may break the mova instruction that
   loads the jump table address, by moving the jump table too far away.
   We fix that problem by never outputting the constant pool between a mova
   and its label.  */

static rtx
find_barrier (from)
     rtx from;
{
  int count_si = 0;
  int count_hi = 0;
  int found_hi = 0;
  int found_si = 0;
  rtx found_barrier = 0;
  rtx found_mova = 0;

  /* For HImode: range is 510, add 4 because pc counts from address of
     second instruction after this one, subtract 2 for the jump instruction
     that we may need to emit before the table.  This gives 512.
     For SImode: range is 1020, add 4 because pc counts from address of
     second instruction after this one, subtract 2 in case pc is 2 byte
     aligned, subtract 2 for the jump instruction that we may need to emit
     before the table.  This gives 1020.  */
  while (from && count_si < 1020 && count_hi < 512)
    {
      int inc = get_attr_length (from);

      if (GET_CODE (from) == BARRIER)
	found_barrier = from;

      if (broken_move (from))
	{
	  rtx src = SET_SRC (PATTERN (from));

	  if (hi_const (src))
	    {
	      found_hi = 1;
	      /* We put the short constants before the long constants, so
		 we must count the length of short constants in the range
		 for the long constants.  */
	      /* ??? This isn't optimal, but is easy to do.  */
	      if (found_si)
		count_si += 2;
	    }
	  else
	    found_si = 1;
	}

      if (GET_CODE (from) == INSN
	  && GET_CODE (PATTERN (from)) == SET
	  && GET_CODE (SET_SRC (PATTERN (from))) == UNSPEC
	  && XINT (SET_SRC (PATTERN (from)), 1) == 1)
	found_mova = from;
      else if (GET_CODE (from) == JUMP_INSN
	       && (GET_CODE (PATTERN (from)) == ADDR_VEC
		   || GET_CODE (PATTERN (from)) == ADDR_DIFF_VEC))
	found_mova = 0;

      if (found_si)
	count_si += inc;
      if (found_hi)
	count_hi += inc;
      from = NEXT_INSN (from);
    }

  /* Insert the constant pool table before the mova instruction, to prevent
     the mova label reference from going out of range.  */
  if (found_mova)
    from = found_mova;

  if (! found_barrier)
    {
      /* We didn't find a barrier in time to dump our stuff,
	 so we'll make one.  */
      rtx label = gen_label_rtx ();

      /* If we exceeded the range, then we must back up over the last
	 instruction we looked at.  Otherwise, we just need to undo the
	 NEXT_INSN at the end of the loop.  */
      if (count_hi > 512 || count_si > 1020)
	from = PREV_INSN (PREV_INSN (from));
      else
	from = PREV_INSN (from);

      /* Walk back to be just before any jump or label.
	 Putting it before a label reduces the number of times the branch
	 around the constant pool table will be hit.  Putting it before
	 a jump makes it more likely that the bra delay slot will be
	 filled.  */
      while (GET_CODE (from) == JUMP_INSN || GET_CODE (from) == NOTE
	     || GET_CODE (from) == CODE_LABEL)
	from = PREV_INSN (from);

      from = emit_jump_insn_after (gen_jump (label), from);
      JUMP_LABEL (from) = label;
      LABEL_NUSES (label) = 1;
      found_barrier = emit_barrier_after (from);
      emit_label_after (label, found_barrier);
    }

  return found_barrier;
}

/* Exported to toplev.c.

   Scan the function looking for move instructions which have to be changed to
   pc-relative loads and insert the literal tables.  */

void
machine_dependent_reorg (first)
     rtx first;
{
  rtx insn;

  for (insn = first; insn; insn = NEXT_INSN (insn))
    {
      if (broken_move (insn))
	{
	  rtx scan;
	  /* Scan ahead looking for a barrier to stick the constant table
	     behind.  */
	  rtx barrier = find_barrier (insn);

	  /* Now find all the moves between the points and modify them.  */
	  for (scan = insn; scan != barrier; scan = NEXT_INSN (scan))
	    {
	      if (broken_move (scan))
		{
		  rtx pat = PATTERN (scan);
		  rtx src = SET_SRC (pat);
		  rtx dst = SET_DEST (pat);
		  enum machine_mode mode = GET_MODE (dst);
		  rtx lab;
		  rtx newinsn;
		  rtx newsrc;

		  if (mode == SImode && hi_const (src))
		    {
		      int offset = 0;

		      mode = HImode;
		      while (GET_CODE (dst) == SUBREG)
			{
			  offset += SUBREG_WORD (dst);
			  dst = SUBREG_REG (dst);
			}
		      dst = gen_rtx (REG, HImode, REGNO (dst) + offset);
		    }

		  lab = add_constant (src, mode);
		  newsrc = gen_rtx (MEM, mode,
				    gen_rtx (LABEL_REF, VOIDmode, lab));
		  RTX_UNCHANGING_P (newsrc) = 1;
		  newinsn = emit_insn_after (gen_rtx (SET, VOIDmode,
						      dst, newsrc), scan);

		  delete_insn (scan);
		  scan = newinsn;
		}
	    }
	  dump_table (barrier);
	}
    }
}

/* Dump out instruction addresses, which is useful for debugging the
   constant pool table stuff.  */

/* ??? This is unnecessary, and probably should be deleted.  This makes
   the insn_addresses declaration above unnecessary.  */

/* ??? The addresses printed by this routine for insns are nonsense for
   insns which are inside of a sequence where none of the inner insns have
   variable length.  This is because the second pass of shorten_branches
   does not bother to update them.  */

void
final_prescan_insn (insn, opvec, noperands)
     rtx insn;
     rtx *opvec;
     int noperands;
{
  if (TARGET_DUMPISIZE)
    fprintf (asm_out_file, "\n! at %04x\n", insn_addresses[INSN_UID (insn)]);
}

/* Dump out any constants accumulated in the final pass.  These will
   will only be labels.  */

char *
output_jump_label_table ()
{
  int i;

  if (pool_size)
    {
      fprintf (asm_out_file, "\t.align 2\n");
      for (i = 0; i < pool_size; i++)
	{
	  pool_node *p = &pool_vector[i];

	  ASM_OUTPUT_INTERNAL_LABEL (asm_out_file, "L",
				     CODE_LABEL_NUMBER (p->label));
	  output_asm_insn (".long	%O0", &p->value);
	}
      pool_size = 0;
    }

  return "";
}

/* A full frame looks like:

   arg-5
   arg-4
   [ if current_function_anonymous_args
   arg-3
   arg-2
   arg-1
   arg-0 ]
   saved-fp
   saved-r10
   saved-r11
   saved-r12
   saved-pr
   local-n
   ..
   local-1
   local-0        <- fp points here.  */

/* Number of bytes pushed for anonymous args, used to pass information
   between expand_prologue and expand_epilogue.  */

static int extra_push;

/* Adjust the stack and return the number of bytes taken to do it.  */

static void
output_stack_adjust (size, reg)
     int size;
     rtx reg;
{
  if (size)
    {
      rtx val = GEN_INT (size);
      rtx insn;

      if (! CONST_OK_FOR_I (size))
	{
	  rtx reg = gen_rtx (REG, SImode, 3);
	  emit_insn (gen_movsi (reg, val));
	  val = reg;
	}

      insn = gen_addsi3 (reg, reg, val);
      emit_insn (insn);
    }
}

/* Output RTL to push register RN onto the stack.  */

static void
push (rn)
     int rn;
{
  rtx x;
  x = emit_insn (gen_push (gen_rtx (REG, SImode, rn)));
  REG_NOTES (x) = gen_rtx (EXPR_LIST, REG_INC,
			   gen_rtx(REG, SImode, STACK_POINTER_REGNUM), 0);
}

/* Output RTL to pop register RN from the stack.  */

static void
pop (rn)