gcc/optabs.c

/* Expand the basic unary and binary arithmetic operations, for GNU compiler.
   Copyright (C) 1987-2019 Free Software Foundation, Inc.

This file is part of GCC.

GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
Software Foundation; either version 3, or (at your option) any later
version.

GCC is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or
FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
for more details.

You should have received a copy of the GNU General Public License
along with GCC; see the file COPYING3.  If not see
<http://www.gnu.org/licenses/>.  */


#include "config.h"
#include "system.h"
#include "coretypes.h"
#include "backend.h"
#include "target.h"
#include "rtl.h"
#include "tree.h"
#include "memmodel.h"
#include "predict.h"
#include "tm_p.h"
#include "expmed.h"
#include "optabs.h"
#include "emit-rtl.h"
#include "recog.h"
#include "diagnostic-core.h"
#include "rtx-vector-builder.h"

/* Include insn-config.h before expr.h so that HAVE_conditional_move
   is properly defined.  */
#include "stor-layout.h"
#include "except.h"
#include "dojump.h"
#include "explow.h"
#include "expr.h"
#include "optabs-tree.h"
#include "libfuncs.h"

static void prepare_float_lib_cmp (rtx, rtx, enum rtx_code, rtx *,
				   machine_mode *);
static rtx expand_unop_direct (machine_mode, optab, rtx, rtx, int);
static void emit_libcall_block_1 (rtx_insn *, rtx, rtx, rtx, bool);

/* Debug facility for use in GDB.  */
void debug_optab_libfuncs (void);

/* Add a REG_EQUAL note to the last insn in INSNS.  TARGET is being set to
   the result of operation CODE applied to OP0 (and OP1 if it is a binary
   operation).  OP0_MODE is OP0's mode.

   If the last insn does not set TARGET, don't do anything, but return 1.

   If the last insn or a previous insn sets TARGET and TARGET is one of OP0
   or OP1, don't add the REG_EQUAL note but return 0.  Our caller can then
   try again, ensuring that TARGET is not one of the operands.  */

static int
add_equal_note (rtx_insn *insns, rtx target, enum rtx_code code, rtx op0,
		rtx op1, machine_mode op0_mode)
{
  rtx_insn *last_insn;
  rtx set;
  rtx note;

  gcc_assert (insns && INSN_P (insns) && NEXT_INSN (insns));

  if (GET_RTX_CLASS (code) != RTX_COMM_ARITH
      && GET_RTX_CLASS (code) != RTX_BIN_ARITH
      && GET_RTX_CLASS (code) != RTX_COMM_COMPARE
      && GET_RTX_CLASS (code) != RTX_COMPARE
      && GET_RTX_CLASS (code) != RTX_UNARY)
    return 1;

  if (GET_CODE (target) == ZERO_EXTRACT)
    return 1;

  for (last_insn = insns;
       NEXT_INSN (last_insn) != NULL_RTX;
       last_insn = NEXT_INSN (last_insn))
    ;

  /* If TARGET is in OP0 or OP1, punt.  We'd end up with a note referencing
     a value changing in the insn, so the note would be invalid for CSE.  */
  if (reg_overlap_mentioned_p (target, op0)
      || (op1 && reg_overlap_mentioned_p (target, op1)))
    {
      if (MEM_P (target)
	  && (rtx_equal_p (target, op0)
	      || (op1 && rtx_equal_p (target, op1))))
	{
	  /* For MEM target, with MEM = MEM op X, prefer no REG_EQUAL note
	     over expanding it as temp = MEM op X, MEM = temp.  If the target
	     supports MEM = MEM op X instructions, it is sometimes too hard
	     to reconstruct that form later, especially if X is also a memory,
	     and due to multiple occurrences of addresses the address might
	     be forced into register unnecessarily.
	     Note that not emitting the REG_EQUIV note might inhibit
	     CSE in some cases.  */
	  set = single_set (last_insn);
	  if (set
	      && GET_CODE (SET_SRC (set)) == code
	      && MEM_P (SET_DEST (set))
	      && (rtx_equal_p (SET_DEST (set), XEXP (SET_SRC (set), 0))
		  || (op1 && rtx_equal_p (SET_DEST (set),
					  XEXP (SET_SRC (set), 1)))))
	    return 1;
	}
      return 0;
    }

  set = set_for_reg_notes (last_insn);
  if (set == NULL_RTX)
    return 1;

  if (! rtx_equal_p (SET_DEST (set), target)
      /* For a STRICT_LOW_PART, the REG_NOTE applies to what is inside it.  */
      && (GET_CODE (SET_DEST (set)) != STRICT_LOW_PART
	  || ! rtx_equal_p (XEXP (SET_DEST (set), 0), target)))
    return 1;

  if (GET_RTX_CLASS (code) == RTX_UNARY)
    switch (code)
      {
      case FFS:
      case CLZ:
      case CTZ:
      case CLRSB:
      case POPCOUNT:
      case PARITY:
      case BSWAP:
	if (op0_mode != VOIDmode && GET_MODE (target) != op0_mode)
	  {
	    note = gen_rtx_fmt_e (code, op0_mode, copy_rtx (op0));
	    if (GET_MODE_UNIT_SIZE (op0_mode)
		> GET_MODE_UNIT_SIZE (GET_MODE (target)))
	      note = simplify_gen_unary (TRUNCATE, GET_MODE (target),
					 note, op0_mode);
	    else
	      note = simplify_gen_unary (ZERO_EXTEND, GET_MODE (target),
					 note, op0_mode);
	    break;
	  }
	/* FALLTHRU */
      default:
	note = gen_rtx_fmt_e (code, GET_MODE (target), copy_rtx (op0));
	break;
      }
  else
    note = gen_rtx_fmt_ee (code, GET_MODE (target), copy_rtx (op0), copy_rtx (op1));

  set_unique_reg_note (last_insn, REG_EQUAL, note);

  return 1;
}

/* Given two input operands, OP0 and OP1, determine what the correct from_mode
   for a widening operation would be.  In most cases this would be OP0, but if
   that's a constant it'll be VOIDmode, which isn't useful.  */

static machine_mode
widened_mode (machine_mode to_mode, rtx op0, rtx op1)
{
  machine_mode m0 = GET_MODE (op0);
  machine_mode m1 = GET_MODE (op1);
  machine_mode result;

  if (m0 == VOIDmode && m1 == VOIDmode)
    return to_mode;
  else if (m0 == VOIDmode || GET_MODE_UNIT_SIZE (m0) < GET_MODE_UNIT_SIZE (m1))
    result = m1;
  else
    result = m0;

  if (GET_MODE_UNIT_SIZE (result) > GET_MODE_UNIT_SIZE (to_mode))
    return to_mode;

  return result;
}

/* Widen OP to MODE and return the rtx for the widened operand.  UNSIGNEDP
   says whether OP is signed or unsigned.  NO_EXTEND is nonzero if we need
   not actually do a sign-extend or zero-extend, but can leave the
   higher-order bits of the result rtx undefined, for example, in the case
   of logical operations, but not right shifts.  */

static rtx
widen_operand (rtx op, machine_mode mode, machine_mode oldmode,
	       int unsignedp, int no_extend)
{
  rtx result;
  scalar_int_mode int_mode;

  /* If we don't have to extend and this is a constant, return it.  */
  if (no_extend && GET_MODE (op) == VOIDmode)
    return op;

  /* If we must extend do so.  If OP is a SUBREG for a promoted object, also
     extend since it will be more efficient to do so unless the signedness of
     a promoted object differs from our extension.  */
  if (! no_extend
      || !is_a <scalar_int_mode> (mode, &int_mode)
      || (GET_CODE (op) == SUBREG && SUBREG_PROMOTED_VAR_P (op)
	  && SUBREG_CHECK_PROMOTED_SIGN (op, unsignedp)))
    return convert_modes (mode, oldmode, op, unsignedp);

  /* If MODE is no wider than a single word, we return a lowpart or paradoxical
     SUBREG.  */
  if (GET_MODE_SIZE (int_mode) <= UNITS_PER_WORD)
    return gen_lowpart (int_mode, force_reg (GET_MODE (op), op));

  /* Otherwise, get an object of MODE, clobber it, and set the low-order
     part to OP.  */

  result = gen_reg_rtx (int_mode);
  emit_clobber (result);
  emit_move_insn (gen_lowpart (GET_MODE (op), result), op);
  return result;
}

/* Expand vector widening operations.

   There are two different classes of operations handled here:
   1) Operations whose result is wider than all the arguments to the operation.
      Examples: VEC_UNPACK_HI/LO_EXPR, VEC_WIDEN_MULT_HI/LO_EXPR
      In this case OP0 and optionally OP1 would be initialized,
      but WIDE_OP wouldn't (not relevant for this case).
   2) Operations whose result is of the same size as the last argument to the
      operation, but wider than all the other arguments to the operation.
      Examples: WIDEN_SUM_EXPR, VEC_DOT_PROD_EXPR.
      In the case WIDE_OP, OP0 and optionally OP1 would be initialized.

   E.g, when called to expand the following operations, this is how
   the arguments will be initialized:
                                nops    OP0     OP1     WIDE_OP
   widening-sum                 2       oprnd0  -       oprnd1
   widening-dot-product         3       oprnd0  oprnd1  oprnd2
   widening-mult                2       oprnd0  oprnd1  -
   type-promotion (vec-unpack)  1       oprnd0  -       -  */

rtx
expand_widen_pattern_expr (sepops ops, rtx op0, rtx op1, rtx wide_op,
			   rtx target, int unsignedp)
{
  struct expand_operand eops[4];
  tree oprnd0, oprnd1, oprnd2;
  machine_mode wmode = VOIDmode, tmode0, tmode1 = VOIDmode;
  optab widen_pattern_optab;
  enum insn_code icode;
  int nops = TREE_CODE_LENGTH (ops->code);
  int op;
  bool sbool = false;

  oprnd0 = ops->op0;
  tmode0 = TYPE_MODE (TREE_TYPE (oprnd0));
  if (ops->code == VEC_UNPACK_FIX_TRUNC_HI_EXPR
      || ops->code == VEC_UNPACK_FIX_TRUNC_LO_EXPR)
    /* The sign is from the result type rather than operand's type
       for these ops.  */
    widen_pattern_optab
      = optab_for_tree_code (ops->code, ops->type, optab_default);
  else if ((ops->code == VEC_UNPACK_HI_EXPR
	    || ops->code == VEC_UNPACK_LO_EXPR)
	   && VECTOR_BOOLEAN_TYPE_P (ops->type)
	   && VECTOR_BOOLEAN_TYPE_P (TREE_TYPE (oprnd0))
	   && TYPE_MODE (ops->type) == TYPE_MODE (TREE_TYPE (oprnd0))
	   && SCALAR_INT_MODE_P (TYPE_MODE (ops->type)))
    {
      /* For VEC_UNPACK_{LO,HI}_EXPR if the mode of op0 and result is
	 the same scalar mode for VECTOR_BOOLEAN_TYPE_P vectors, use
	 vec_unpacks_sbool_{lo,hi}_optab, so that we can pass in
	 the pattern number of elements in the wider vector.  */
      widen_pattern_optab
	= (ops->code == VEC_UNPACK_HI_EXPR
	   ? vec_unpacks_sbool_hi_optab : vec_unpacks_sbool_lo_optab);
      sbool = true;
    }
  else
    widen_pattern_optab
      = optab_for_tree_code (ops->code, TREE_TYPE (oprnd0), optab_default);
  if (ops->code == WIDEN_MULT_PLUS_EXPR
      || ops->code == WIDEN_MULT_MINUS_EXPR)
    icode = find_widening_optab_handler (widen_pattern_optab,
					 TYPE_MODE (TREE_TYPE (ops->op2)),
					 tmode0);
  else
    icode = optab_handler (widen_pattern_optab, tmode0);
  gcc_assert (icode != CODE_FOR_nothing);

  if (nops >= 2)
    {
      oprnd1 = ops->op1;
      tmode1 = TYPE_MODE (TREE_TYPE (oprnd1));
    }
  else if (sbool)
    {
      nops = 2;
      op1 = GEN_INT (TYPE_VECTOR_SUBPARTS (TREE_TYPE (oprnd0)).to_constant ());
      tmode1 = tmode0;
    }

  /* The last operand is of a wider mode than the rest of the operands.  */
  if (nops == 2)
    wmode = tmode1;
  else if (nops == 3)
    {
      gcc_assert (tmode1 == tmode0);
      gcc_assert (op1);
      oprnd2 = ops->op2;
      wmode = TYPE_MODE (TREE_TYPE (oprnd2));
    }

  op = 0;
  create_output_operand (&eops[op++], target, TYPE_MODE (ops->type));
  create_convert_operand_from (&eops[op++], op0, tmode0, unsignedp);
  if (op1)
    create_convert_operand_from (&eops[op++], op1, tmode1, unsignedp);
  if (wide_op)
    create_convert_operand_from (&eops[op++], wide_op, wmode, unsignedp);
  expand_insn (icode, op, eops);
  return eops[0].value;
}

/* Generate code to perform an operation specified by TERNARY_OPTAB
   on operands OP0, OP1 and OP2, with result having machine-mode MODE.

   UNSIGNEDP is for the case where we have to widen the operands
   to perform the operation.  It says to use zero-extension.

   If TARGET is nonzero, the value
   is generated there, if it is convenient to do so.
   In all cases an rtx is returned for the locus of the value;
   this may or may not be TARGET.  */

rtx
expand_ternary_op (machine_mode mode, optab ternary_optab, rtx op0,
		   rtx op1, rtx op2, rtx target, int unsignedp)
{
  struct expand_operand ops[4];
  enum insn_code icode = optab_handler (ternary_optab, mode);

  gcc_assert (optab_handler (ternary_optab, mode) != CODE_FOR_nothing);

  create_output_operand (&ops[0], target, mode);
  create_convert_operand_from (&ops[1], op0, mode, unsignedp);
  create_convert_operand_from (&ops[2], op1, mode, unsignedp);
  create_convert_operand_from (&ops[3], op2, mode, unsignedp);
  expand_insn (icode, 4, ops);
  return ops[0].value;
}


/* Like expand_binop, but return a constant rtx if the result can be
   calculated at compile time.  The arguments and return value are
   otherwise the same as for expand_binop.  */

rtx
simplify_expand_binop (machine_mode mode, optab binoptab,
		       rtx op0, rtx op1, rtx target, int unsignedp,
		       enum optab_methods methods)
{
  if (CONSTANT_P (op0) && CONSTANT_P (op1))
    {
      rtx x = simplify_binary_operation (optab_to_code (binoptab),
					 mode, op0, op1);
      if (x)
	return x;
    }

  return expand_binop (mode, binoptab, op0, op1, target, unsignedp, methods);
}

/* Like simplify_expand_binop, but always put the result in TARGET.
   Return true if the expansion succeeded.  */

bool
force_expand_binop (machine_mode mode, optab binoptab,
		    rtx op0, rtx op1, rtx target, int unsignedp,
		    enum optab_methods methods)
{
  rtx x = simplify_expand_binop (mode, binoptab, op0, op1,
				 target, unsignedp, methods);
  if (x == 0)
    return false;
  if (x != target)
    emit_move_insn (target, x);
  return true;
}

/* Create a new vector value in VMODE with all elements set to OP.  The
   mode of OP must be the element mode of VMODE.  If OP is a constant,
   then the return value will be a constant.  */

rtx
expand_vector_broadcast (machine_mode vmode, rtx op)
{
  int n;
  rtvec vec;

  gcc_checking_assert (VECTOR_MODE_P (vmode));

  if (valid_for_const_vector_p (vmode, op))
    return gen_const_vec_duplicate (vmode, op);

  insn_code icode = optab_handler (vec_duplicate_optab, vmode);
  if (icode != CODE_FOR_nothing)
    {
      struct expand_operand ops[2];
      create_output_operand (&ops[0], NULL_RTX, vmode);
      create_input_operand (&ops[1], op, GET_MODE (op));
      expand_insn (icode, 2, ops);
      return ops[0].value;
    }

  if (!GET_MODE_NUNITS (vmode).is_constant (&n))
    return NULL;

  /* ??? If the target doesn't have a vec_init, then we have no easy way
     of performing this operation.  Most of this sort of generic support
     is hidden away in the vector lowering support in gimple.  */
  icode = convert_optab_handler (vec_init_optab, vmode,
				 GET_MODE_INNER (vmode));
  if (icode == CODE_FOR_nothing)
    return NULL;

  vec = rtvec_alloc (n);
  for (int i = 0; i < n; ++i)
    RTVEC_ELT (vec, i) = op;
  rtx ret = gen_reg_rtx (vmode);
  emit_insn (GEN_FCN (icode) (ret, gen_rtx_PARALLEL (vmode, vec)));

  return ret;
}

/* This subroutine of expand_doubleword_shift handles the cases in which
   the effective shift value is >= BITS_PER_WORD.  The arguments and return
   value are the same as for the parent routine, except that SUPERWORD_OP1
   is the shift count to use when shifting OUTOF_INPUT into INTO_TARGET.
   INTO_TARGET may be null if the caller has decided to calculate it.  */

static bool
expand_superword_shift (optab binoptab, rtx outof_input, rtx superword_op1,
			rtx outof_target, rtx into_target,
			int unsignedp, enum optab_methods methods)
{
  if (into_target != 0)
    if (!force_expand_binop (word_mode, binoptab, outof_input, superword_op1,
			     into_target, unsignedp, methods))
      return false;

  if (outof_target != 0)
    {
      /* For a signed right shift, we must fill OUTOF_TARGET with copies
	 of the sign bit, otherwise we must fill it with zeros.  */
      if (binoptab != ashr_optab)
	emit_move_insn (outof_target, CONST0_RTX (word_mode));
      else
	if (!force_expand_binop (word_mode, binoptab, outof_input,
				 gen_int_shift_amount (word_mode,
						       BITS_PER_WORD - 1),
				 outof_target, unsignedp, methods))
	  return false;
    }
  return true;
}

/* This subroutine of expand_doubleword_shift handles the cases in which
   the effective shift value is < BITS_PER_WORD.  The arguments and return
   value are the same as for the parent routine.  */

static bool
expand_subword_shift (scalar_int_mode op1_mode, optab binoptab,
		      rtx outof_input, rtx into_input, rtx op1,
		      rtx outof_target, rtx into_target,
		      int unsignedp, enum optab_methods methods,
		      unsigned HOST_WIDE_INT shift_mask)
{
  optab reverse_unsigned_shift, unsigned_shift;
  rtx tmp, carries;

  reverse_unsigned_shift = (binoptab == ashl_optab ? lshr_optab : ashl_optab);
  unsigned_shift = (binoptab == ashl_optab ? ashl_optab : lshr_optab);

  /* The low OP1 bits of INTO_TARGET come from the high bits of OUTOF_INPUT.
     We therefore need to shift OUTOF_INPUT by (BITS_PER_WORD - OP1) bits in
     the opposite direction to BINOPTAB.  */
  if (CONSTANT_P (op1) || shift_mask >= BITS_PER_WORD)
    {
      carries = outof_input;
      tmp = immed_wide_int_const (wi::shwi (BITS_PER_WORD,
					    op1_mode), op1_mode);
      tmp = simplify_expand_binop (op1_mode, sub_optab, tmp, op1,
				   0, true, methods);
    }
  else
    {
      /* We must avoid shifting by BITS_PER_WORD bits since that is either
	 the same as a zero shift (if shift_mask == BITS_PER_WORD - 1) or
	 has unknown behavior.  Do a single shift first, then shift by the
	 remainder.  It's OK to use ~OP1 as the remainder if shift counts
	 are truncated to the mode size.  */
      carries = expand_binop (word_mode, reverse_unsigned_shift,
			      outof_input, const1_rtx, 0, unsignedp, methods);
      if (shift_mask == BITS_PER_WORD - 1)
	{
	  tmp = immed_wide_int_const
	    (wi::minus_one (GET_MODE_PRECISION (op1_mode)), op1_mode);
	  tmp = simplify_expand_binop (op1_mode, xor_optab, op1, tmp,
				       0, true, methods);
	}
      else
	{
	  tmp = immed_wide_int_const (wi::shwi (BITS_PER_WORD - 1,
						op1_mode), op1_mode);
	  tmp = simplify_expand_binop (op1_mode, sub_optab, tmp, op1,
				       0, true, methods);
	}
    }
  if (tmp == 0 || carries == 0)
    return false;
  carries = expand_binop (word_mode, reverse_unsigned_shift,
			  carries, tmp, 0, unsignedp, methods);
  if (carries == 0)
    return false;

  /* Shift INTO_INPUT logically by OP1.  This is the last use of INTO_INPUT
     so the result can go directly into INTO_TARGET if convenient.  */
  tmp = expand_binop (word_mode, unsigned_shift, into_input, op1,
		      into_target, unsignedp, methods);
  if (tmp == 0)
    return false;

  /* Now OR in the bits carried over from OUTOF_INPUT.  */
  if (!force_expand_binop (word_mode, ior_optab, tmp, carries,
			   into_target, unsignedp, methods))
    return false;

  /* Use a standard word_mode shift for the out-of half.  */
  if (outof_target != 0)
    if (!force_expand_binop (word_mode, binoptab, outof_input, op1,
			     outof_target, unsignedp, methods))
      return false;

  return true;
}


/* Try implementing expand_doubleword_shift using conditional moves.
   The shift is by < BITS_PER_WORD if (CMP_CODE CMP1 CMP2) is true,
   otherwise it is by >= BITS_PER_WORD.  SUBWORD_OP1 and SUPERWORD_OP1
   are the shift counts to use in the former and latter case.  All other
   arguments are the same as the parent routine.  */

static bool
expand_doubleword_shift_condmove (scalar_int_mode op1_mode, optab binoptab,
				  enum rtx_code cmp_code, rtx cmp1, rtx cmp2,
				  rtx outof_input, rtx into_input,
				  rtx subword_op1, rtx superword_op1,
				  rtx outof_target, rtx into_target,
				  int unsignedp, enum optab_methods methods,
				  unsigned HOST_WIDE_INT shift_mask)
{
  rtx outof_superword, into_superword;

  /* Put the superword version of the output into OUTOF_SUPERWORD and
     INTO_SUPERWORD.  */
  outof_superword = outof_target != 0 ? gen_reg_rtx (word_mode) : 0;
  if (outof_target != 0 && subword_op1 == superword_op1)
    {
      /* The value INTO_TARGET >> SUBWORD_OP1, which we later store in
	 OUTOF_TARGET, is the same as the value of INTO_SUPERWORD.  */
      into_superword = outof_target;
      if (!expand_superword_shift (binoptab, outof_input, superword_op1,
				   outof_superword, 0, unsignedp, methods))
	return false;
    }
  else
    {
      into_superword = gen_reg_rtx (word_mode);
      if (!expand_superword_shift (binoptab, outof_input, superword_op1,
				   outof_superword, into_superword,
				   unsignedp, methods))
	return false;
    }

  /* Put the subword version directly in OUTOF_TARGET and INTO_TARGET.  */
  if (!expand_subword_shift (op1_mode, binoptab,
			     outof_input, into_input, subword_op1,
			     outof_target, into_target,
			     unsignedp, methods, shift_mask))
    return false;

  /* Select between them.  Do the INTO half first because INTO_SUPERWORD
     might be the current value of OUTOF_TARGET.  */
  if (!emit_conditional_move (into_target, cmp_code, cmp1, cmp2, op1_mode,
			      into_target, into_superword, word_mode, false))
    return false;

  if (outof_target != 0)
    if (!emit_conditional_move (outof_target, cmp_code, cmp1, cmp2, op1_mode,
				outof_target, outof_superword,
				word_mode, false))
      return false;

  return true;
}

/* Expand a doubleword shift (ashl, ashr or lshr) using word-mode shifts.
   OUTOF_INPUT and INTO_INPUT are the two word-sized halves of the first
   input operand; the shift moves bits in the direction OUTOF_INPUT->
   INTO_TARGET.  OUTOF_TARGET and INTO_TARGET are the equivalent words
   of the target.  OP1 is the shift count and OP1_MODE is its mode.
   If OP1 is constant, it will have been truncated as appropriate
   and is known to be nonzero.

   If SHIFT_MASK is zero, the result of word shifts is undefined when the
   shift count is outside the range [0, BITS_PER_WORD).  This routine must
   avoid generating such shifts for OP1s in the range [0, BITS_PER_WORD * 2).

   If SHIFT_MASK is nonzero, all word-mode shift counts are effectively
   masked by it and shifts in the range [BITS_PER_WORD, SHIFT_MASK) will
   fill with zeros or sign bits as appropriate.

   If SHIFT_MASK is BITS_PER_WORD - 1, this routine will synthesize
   a doubleword shift whose equivalent mask is BITS_PER_WORD * 2 - 1.
   Doing this preserves semantics required by SHIFT_COUNT_TRUNCATED.
   In all other cases, shifts by values outside [0, BITS_PER_UNIT * 2)
   are undefined.

   BINOPTAB, UNSIGNEDP and METHODS are as for expand_binop.  This function
   may not use INTO_INPUT after modifying INTO_TARGET, and similarly for
   OUTOF_INPUT and OUTOF_TARGET.  OUTOF_TARGET can be null if the parent
   function wants to calculate it itself.

   Return true if the shift could be successfully synthesized.  */

static bool
expand_doubleword_shift (scalar_int_mode op1_mode, optab binoptab,
			 rtx outof_input, rtx into_input, rtx op1,
			 rtx outof_target, rtx into_target,
			 int unsignedp, enum optab_methods methods,
			 unsigned HOST_WIDE_INT shift_mask)
{
  rtx superword_op1, tmp, cmp1, cmp2;
  enum rtx_code cmp_code;

  /* See if word-mode shifts by BITS_PER_WORD...BITS_PER_WORD * 2 - 1 will
     fill the result with sign or zero bits as appropriate.  If so, the value
     of OUTOF_TARGET will always be (SHIFT OUTOF_INPUT OP1).   Recursively call
     this routine to calculate INTO_TARGET (which depends on both OUTOF_INPUT
     and INTO_INPUT), then emit code to set up OUTOF_TARGET.

     This isn't worthwhile for constant shifts since the optimizers will
     cope better with in-range shift counts.  */
  if (shift_mask >= BITS_PER_WORD
      && outof_target != 0
      && !CONSTANT_P (op1))
    {
      if (!expand_doubleword_shift (op1_mode, binoptab,
				    outof_input, into_input, op1,
				    0, into_target,
				    unsignedp, methods, shift_mask))
	return false;
      if (!force_expand_binop (word_mode, binoptab, outof_input, op1,
			       outof_target, unsignedp, methods))
	return false;
      return true;
    }

  /* Set CMP_CODE, CMP1 and CMP2 so that the rtx (CMP_CODE CMP1 CMP2)
     is true when the effective shift value is less than BITS_PER_WORD.
     Set SUPERWORD_OP1 to the shift count that should be used to shift
     OUTOF_INPUT into INTO_TARGET when the condition is false.  */
  tmp = immed_wide_int_const (wi::shwi (BITS_PER_WORD, op1_mode), op1_mode);
  if (!CONSTANT_P (op1) && shift_mask == BITS_PER_WORD - 1)
    {
      /* Set CMP1 to OP1 & BITS_PER_WORD.  The result is zero iff OP1
	 is a subword shift count.  */
      cmp1 = simplify_expand_binop (op1_mode, and_optab, op1, tmp,
				    0, true, methods);
      cmp2 = CONST0_RTX (op1_mode);
      cmp_code = EQ;
      superword_op1 = op1;
    }
  else
    {
      /* Set CMP1 to OP1 - BITS_PER_WORD.  */
      cmp1 = simplify_expand_binop (op1_mode, sub_optab, op1, tmp,
				    0, true, methods);
      cmp2 = CONST0_RTX (op1_mode);
      cmp_code = LT;
      superword_op1 = cmp1;
    }
  if (cmp1 == 0)
    return false;

  /* If we can compute the condition at compile time, pick the
     appropriate subroutine.  */
  tmp = simplify_relational_operation (cmp_code, SImode, op1_mode, cmp1, cmp2);
  if (tmp != 0 && CONST_INT_P (tmp))
    {
      if (tmp == const0_rtx)
	return expand_superword_shift (binoptab, outof_input, superword_op1,
				       outof_target, into_target,
				       unsignedp, methods);
      else
	return expand_subword_shift (op1_mode, binoptab,
				     outof_input, into_input, op1,
				     outof_target, into_target,
				     unsignedp, methods, shift_mask);
    }

  /* Try using conditional moves to generate straight-line code.  */
  if (HAVE_conditional_move)
    {
      rtx_insn *start = get_last_insn ();
      if (expand_doubleword_shift_condmove (op1_mode, binoptab,
					    cmp_code, cmp1, cmp2,
					    outof_input, into_input,
					    op1, superword_op1,
					    outof_target, into_target,
					    unsignedp, methods, shift_mask))
	return true;
      delete_insns_since (start);
    }

  /* As a last resort, use branches to select the correct alternative.  */
  rtx_code_label *subword_label = gen_label_rtx ();
  rtx_code_label *done_label = gen_label_rtx ();

  NO_DEFER_POP;
  do_compare_rtx_and_jump (cmp1, cmp2, cmp_code, false, op1_mode,
			   0, 0, subword_label,
			   profile_probability::uninitialized ());
  OK_DEFER_POP;

  if (!expand_superword_shift (binoptab, outof_input, superword_op1,
			       outof_target, into_target,
			       unsignedp, methods))
    return false;

  emit_jump_insn (targetm.gen_jump (done_label));
  emit_barrier ();
  emit_label (subword_label);

  if (!expand_subword_shift (op1_mode, binoptab,
			     outof_input, into_input, op1,
			     outof_target, into_target,
			     unsignedp, methods, shift_mask))
    return false;

  emit_label (done_label);
  return true;
}

/* Subroutine of expand_binop.  Perform a double word multiplication of
   operands OP0 and OP1 both of mode MODE, which is exactly twice as wide
   as the target's word_mode.  This function return NULL_RTX if anything
   goes wrong, in which case it may have already emitted instructions
   which need to be deleted.

   If we want to multiply two two-word values and have normal and widening
   multiplies of single-word values, we can do this with three smaller
   multiplications.

   The multiplication proceeds as follows:
			         _______________________
			        [__op0_high_|__op0_low__]
			         _______________________
        *			[__op1_high_|__op1_low__]
        _______________________________________________
			         _______________________
    (1)				[__op0_low__*__op1_low__]
		     _______________________
    (2a)	    [__op0_low__*__op1_high_]
		     _______________________
    (2b)	    [__op0_high_*__op1_low__]
         _______________________
    (3) [__op0_high_*__op1_high_]


  This gives a 4-word result.  Since we are only interested in the
  lower 2 words, partial result (3) and the upper words of (2a) and
  (2b) don't need to be calculated.  Hence (2a) and (2b) can be
  calculated using non-widening multiplication.

  (1), however, needs to be calculated with an unsigned widening
  multiplication.  If this operation is not directly supported we
  try using a signed widening multiplication and adjust the result.
  This adjustment works as follows:

      If both operands are positive then no adjustment is needed.

      If the operands have different signs, for example op0_low < 0 and
      op1_low >= 0, the instruction treats the most significant bit of
      op0_low as a sign bit instead of a bit with significance
      2**(BITS_PER_WORD-1), i.e. the instruction multiplies op1_low
      with 2**BITS_PER_WORD - op0_low, and two's complements the
      result.  Conclusion: We need to add op1_low * 2**BITS_PER_WORD to
      the result.

      Similarly, if both operands are negative, we need to add
      (op0_low + op1_low) * 2**BITS_PER_WORD.

      We use a trick to adjust quickly.  We logically shift op0_low right
      (op1_low) BITS_PER_WORD-1 steps to get 0 or 1, and add this to
      op0_high (op1_high) before it is used to calculate 2b (2a).  If no
      logical shift exists, we do an arithmetic right shift and subtract
      the 0 or -1.  */

static rtx
expand_doubleword_mult (machine_mode mode, rtx op0, rtx op1, rtx target,
		       bool umulp, enum optab_methods methods)
{
  int low = (WORDS_BIG_ENDIAN ? 1 : 0);
  int high = (WORDS_BIG_ENDIAN ? 0 : 1);
  rtx wordm1 = (umulp ? NULL_RTX
		: gen_int_shift_amount (word_mode, BITS_PER_WORD - 1));
  rtx product, adjust, product_high, temp;

  rtx op0_high = operand_subword_force (op0, high, mode);
  rtx op0_low = operand_subword_force (op0, low, mode);
  rtx op1_high = operand_subword_force (op1, high, mode);
  rtx op1_low = operand_subword_force (op1, low, mode);

  /* If we're using an unsigned multiply to directly compute the product
     of the low-order words of the operands and perform any required
     adjustments of the operands, we begin by trying two more multiplications
     and then computing the appropriate sum.

     We have checked above that the required addition is provided.
     Full-word addition will normally always succeed, especially if
     it is provided at all, so we don't worry about its failure.  The
     multiplication may well fail, however, so we do handle that.  */

  if (!umulp)
    {
      /* ??? This could be done with emit_store_flag where available.  */
      temp = expand_binop (word_mode, lshr_optab, op0_low, wordm1,
			   NULL_RTX, 1, methods);
      if (temp)
	op0_high = expand_binop (word_mode, add_optab, op0_high, temp,
				 NULL_RTX, 0, OPTAB_DIRECT);
      else
	{
	  temp = expand_binop (word_mode, ashr_optab, op0_low, wordm1,
			       NULL_RTX, 0, methods);
	  if (!temp)
	    return NULL_RTX;
	  op0_high = expand_binop (word_mode, sub_optab, op0_high, temp,
				   NULL_RTX, 0, OPTAB_DIRECT);
	}

      if (!op0_high)
	return NULL_RTX;
    }

  adjust = expand_binop (word_mode, smul_optab, op0_high, op1_low,
			 NULL_RTX, 0, OPTAB_DIRECT);
  if (!adjust)
    return NULL_RTX;

  /* OP0_HIGH should now be dead.  */

  if (!umulp)
    {
      /* ??? This could be done with emit_store_flag where available.  */
      temp = expand_binop (word_mode, lshr_optab, op1_low, wordm1,
			   NULL_RTX, 1, methods);
      if (temp)
	op1_high = expand_binop (word_mode, add_optab, op1_high, temp,
				 NULL_RTX, 0, OPTAB_DIRECT);
      else
	{
	  temp = expand_binop (word_mode, ashr_optab, op1_low, wordm1,
			       NULL_RTX, 0, methods);
	  if (!temp)
	    return NULL_RTX;
	  op1_high = expand_binop (word_mode, sub_optab, op1_high, temp,
				   NULL_RTX, 0, OPTAB_DIRECT);
	}

      if (!op1_high)
	return NULL_RTX;
    }

  temp = expand_binop (word_mode, smul_optab, op1_high, op0_low,
		       NULL_RTX, 0, OPTAB_DIRECT);
  if (!temp)
    return NULL_RTX;

  /* OP1_HIGH should now be dead.  */

  adjust = expand_binop (word_mode, add_optab, adjust, temp,
			 NULL_RTX, 0, OPTAB_DIRECT);

  if (target && !REG_P (target))
    target = NULL_RTX;

  /* *_widen_optab needs to determine operand mode, make sure at least
     one operand has non-VOID mode.  */
  if (GET_MODE (op0_low) == VOIDmode && GET_MODE (op1_low) == VOIDmode)
    op0_low = force_reg (word_mode, op0_low);

  if (umulp)
    product = expand_binop (mode, umul_widen_optab, op0_low, op1_low,
			    target, 1, OPTAB_DIRECT);
  else
    product = expand_binop (mode, smul_widen_optab, op0_low, op1_low,
			    target, 1, OPTAB_DIRECT);

  if (!product)
    return NULL_RTX;

  product_high = operand_subword (product, high, 1, mode);
  adjust = expand_binop (word_mode, add_optab, product_high, adjust,
			 NULL_RTX, 0, OPTAB_DIRECT);
  emit_move_insn (product_high, adjust);
  return product;
}

/* Wrapper around expand_binop which takes an rtx code to specify
   the operation to perform, not an optab pointer.  All other
   arguments are the same.  */
rtx
expand_simple_binop (machine_mode mode, enum rtx_code code, rtx op0,
		     rtx op1, rtx target, int unsignedp,
		     enum optab_methods methods)
{
  optab binop = code_to_optab (code);
  gcc_assert (binop);

  return expand_binop (mode, binop, op0, op1, target, unsignedp, methods);
}

/* Return whether OP0 and OP1 should be swapped when expanding a commutative
   binop.  Order them according to commutative_operand_precedence and, if
   possible, try to put TARGET or a pseudo first.  */
static bool
swap_commutative_operands_with_target (rtx target, rtx op0, rtx op1)
{
  int op0_prec = commutative_operand_precedence (op0);
  int op1_prec = commutative_operand_precedence (op1);

  if (op0_prec < op1_prec)
    return true;

  if (op0_prec > op1_prec)
    return false;

  /* With equal precedence, both orders are ok, but it is better if the
     first operand is TARGET, or if both TARGET and OP0 are pseudos.  */
  if (target == 0 || REG_P (target))
    return (REG_P (op1) && !REG_P (op0)) || target == op1;
  else
    return rtx_equal_p (op1, target);
}

/* Return true if BINOPTAB implements a shift operation.  */

static bool
shift_optab_p (optab binoptab)
{
  switch (optab_to_code (binoptab))
    {
    case ASHIFT:
    case SS_ASHIFT:
    case US_ASHIFT:
    case ASHIFTRT:
    case LSHIFTRT:
    case ROTATE:
    case ROTATERT:
      return true;

    default:
      return false;
    }
}

/* Return true if BINOPTAB implements a commutative binary operation.  */

static bool
commutative_optab_p (optab binoptab)
{
  return (GET_RTX_CLASS (optab_to_code (binoptab)) == RTX_COMM_ARITH
	  || binoptab == smul_widen_optab
	  || binoptab == umul_widen_optab
	  || binoptab == smul_highpart_optab
	  || binoptab == umul_highpart_optab);
}

/* X is to be used in mode MODE as operand OPN to BINOPTAB.  If we're
   optimizing, and if the operand is a constant that costs more than
   1 instruction, force the constant into a register and return that
   register.  Return X otherwise.  UNSIGNEDP says whether X is unsigned.  */

static rtx
avoid_expensive_constant (machine_mode mode, optab binoptab,
			  int opn, rtx x, bool unsignedp)
{
  bool speed = optimize_insn_for_speed_p ();

  if (mode != VOIDmode
      && optimize
      && CONSTANT_P (x)
      && (rtx_cost (x, mode, optab_to_code (binoptab), opn, speed)
	  > set_src_cost (x, mode, speed)))
    {
      if (CONST_INT_P (x))
	{
	  HOST_WIDE_INT intval = trunc_int_for_mode (INTVAL (x), mode);
	  if (intval != INTVAL (x))
	    x = GEN_INT (intval);
	}
      else
	x = convert_modes (mode, VOIDmode, x, unsignedp);
      x = force_reg (mode, x);
    }
  return x;
}

/* Helper function for expand_binop: handle the case where there
   is an insn ICODE that directly implements the indicated operation.
   Returns null if this is not possible.  */
static rtx
expand_binop_directly (enum insn_code icode, machine_mode mode, optab binoptab,
		       rtx op0, rtx op1,
		       rtx target, int unsignedp, enum optab_methods methods,
		       rtx_insn *last)
{
  machine_mode xmode0 = insn_data[(int) icode].operand[1].mode;
  machine_mode xmode1 = insn_data[(int) icode].operand[2].mode;
  machine_mode mode0, mode1, tmp_mode;
  struct expand_operand ops[3];
  bool commutative_p;
  rtx_insn *pat;
  rtx xop0 = op0, xop1 = op1;
  bool canonicalize_op1 = false;

  /* If it is a commutative operator and the modes would match
     if we would swap the operands, we can save the conversions.  */
  commutative_p = commutative_optab_p (binoptab);
  if (commutative_p
      && GET_MODE (xop0) != xmode0 && GET_MODE (xop1) != xmode1
      && GET_MODE (xop0) == xmode1 && GET_MODE (xop1) == xmode1)
    std::swap (xop0, xop1);

  /* If we are optimizing, force expensive constants into a register.  */
  xop0 = avoid_expensive_constant (xmode0, binoptab, 0, xop0, unsignedp);
  if (!shift_optab_p (binoptab))
    xop1 = avoid_expensive_constant (xmode1, binoptab, 1, xop1, unsignedp);
  else
    /* Shifts and rotates often use a different mode for op1 from op0;
       for VOIDmode constants we don't know the mode, so force it
       to be canonicalized using convert_modes.  */
    canonicalize_op1 = true;

  /* In case the insn wants input operands in modes different from
     those of the actual operands, convert the operands.  It would
     seem that we don't need to convert CONST_INTs, but we do, so
     that they're properly zero-extended, sign-extended or truncated
     for their mode.  */

  mode0 = GET_MODE (xop0) != VOIDmode ? GET_MODE (xop0) : mode;
  if (xmode0 != VOIDmode && xmode0 != mode0)
    {
      xop0 = convert_modes (xmode0, mode0, xop0, unsignedp);
      mode0 = xmode0;
    }

  mode1 = ((GET_MODE (xop1) != VOIDmode || canonicalize_op1)
	   ? GET_MODE (xop1) : mode);
  if (xmode1 != VOIDmode && xmode1 != mode1)
    {
      xop1 = convert_modes (xmode1, mode1, xop1, unsignedp);
      mode1 = xmode1;
    }

  /* If operation is commutative,
     try to make the first operand a register.
     Even better, try to make it the same as the target.
     Also try to make the last operand a constant.  */
  if (commutative_p
      && swap_commutative_operands_with_target (target, xop0, xop1))
    std::swap (xop0, xop1);

  /* Now, if insn's predicates don't allow our operands, put them into
     pseudo regs.  */

  if (binoptab == vec_pack_trunc_optab
      || binoptab == vec_pack_usat_optab
      || binoptab == vec_pack_ssat_optab
      || binoptab == vec_pack_ufix_trunc_optab
      || binoptab == vec_pack_sfix_trunc_optab
      || binoptab == vec_packu_float_optab
      || binoptab == vec_packs_float_optab)
    {
      /* The mode of the result is different then the mode of the
	 arguments.  */
      tmp_mode = insn_data[(int) icode].operand[0].mode;
      if (VECTOR_MODE_P (mode)
	  && maybe_ne (GET_MODE_NUNITS (tmp_mode), 2 * GET_MODE_NUNITS (mode)))
	{
	  delete_insns_since (last);
	  return NULL_RTX;
	}
    }
  else
    tmp_mode = mode;

  create_output_operand (&ops[0], target, tmp_mode);
  create_input_operand (&ops[1], xop0, mode0);
  create_input_operand (&ops[2], xop1, mode1);
  pat = maybe_gen_insn (icode, 3, ops);
  if (pat)
    {
      /* If PAT is composed of more than one insn, try to add an appropriate
	 REG_EQUAL note to it.  If we can't because TEMP conflicts with an
	 operand, call expand_binop again, this time without a target.  */
      if (INSN_P (pat) && NEXT_INSN (pat) != NULL_RTX
	  && ! add_equal_note (pat, ops[0].value,
			       optab_to_code (binoptab),
			       ops[1].value, ops[2].value, mode0))
	{
	  delete_insns_since (last);
	  return expand_binop (mode, binoptab, op0, op1, NULL_RTX,
			       unsignedp, methods);
	}

      emit_insn (pat);
      return ops[0].value;
    }
  delete_insns_since (last);
  return NULL_RTX;
}

/* Generate code to perform an operation specified by BINOPTAB
   on operands OP0 and OP1, with result having machine-mode MODE.

   UNSIGNEDP is for the case where we have to widen the operands
   to perform the operation.  It says to use zero-extension.

   If TARGET is nonzero, the value
   is generated there, if it is convenient to do so.
   In all cases an rtx is returned for the locus of the value;
   this may or may not be TARGET.  */

rtx
expand_binop (machine_mode mode, optab binoptab, rtx op0, rtx op1,
	      rtx target, int unsignedp, enum optab_methods methods)
{
  enum optab_methods next_methods
    = (methods == OPTAB_LIB || methods == OPTAB_LIB_WIDEN
       ? OPTAB_WIDEN : methods);
  enum mode_class mclass;
  enum insn_code icode;
  machine_mode wider_mode;
  scalar_int_mode int_mode;
  rtx libfunc;
  rtx temp;
  rtx_insn *entry_last = get_last_insn ();
  rtx_insn *last;

  mclass = GET_MODE_CLASS (mode);

  /* If subtracting an integer constant, convert this into an addition of
     the negated constant.  */

  if (binoptab == sub_optab && CONST_INT_P (op1))
    {
      op1 = negate_rtx (mode, op1);
      binoptab = add_optab;
    }
  /* For shifts, constant invalid op1 might be expanded from different
     mode than MODE.  As those are invalid, force them to a register
     to avoid further problems during expansion.  */
  else if (CONST_INT_P (op1)
	   && shift_optab_p (binoptab)
	   && UINTVAL (op1) >= GET_MODE_BITSIZE (GET_MODE_INNER (mode)))
    {
      op1 = gen_int_mode (INTVAL (op1), GET_MODE_INNER (mode));
      op1 = force_reg (GET_MODE_INNER (mode), op1);
    }

  /* Record where to delete back to if we backtrack.  */
  last = get_last_insn ();

  /* If we can do it with a three-operand insn, do so.  */

  if (methods != OPTAB_MUST_WIDEN)
    {
      if (convert_optab_p (binoptab))
	{
	  machine_mode from_mode = widened_mode (mode, op0, op1);
	  icode = find_widening_optab_handler (binoptab, mode, from_mode);
	}
      else
	icode = optab_handler (binoptab, mode);
      if (icode != CODE_FOR_nothing)
	{
	  temp = expand_binop_directly (icode, mode, binoptab, op0, op1,
					target, unsignedp, methods, last);
	  if (temp)
	    return temp;
	}
    }

  /* If we were trying to rotate, and that didn't work, try rotating
     the other direction before falling back to shifts and bitwise-or.  */
  if (((binoptab == rotl_optab
	&& (icode = optab_handler (rotr_optab, mode)) != CODE_FOR_nothing)
       || (binoptab == rotr_optab
	   && (icode = optab_handler (rotl_optab, mode)) != CODE_FOR_nothing))
      && is_int_mode (mode, &int_mode))
    {
      optab otheroptab = (binoptab == rotl_optab ? rotr_optab : rotl_optab);
      rtx newop1;
      unsigned int bits = GET_MODE_PRECISION (int_mode);

      if (CONST_INT_P (op1))
	newop1 = gen_int_shift_amount (int_mode, bits - INTVAL (op1));
      else if (targetm.shift_truncation_mask (int_mode) == bits - 1)
        newop1 = negate_rtx (GET_MODE (op1), op1);
      else
        newop1 = expand_binop (GET_MODE (op1), sub_optab,
			       gen_int_mode (bits, GET_MODE (op1)), op1,
			       NULL_RTX, unsignedp, OPTAB_DIRECT);

      temp = expand_binop_directly (icode, int_mode, otheroptab, op0, newop1,
				    target, unsignedp, methods, last);
      if (temp)
	return temp;
    }

  /* If this is a multiply, see if we can do a widening operation that
     takes operands of this mode and makes a wider mode.  */

  if (binoptab == smul_optab
      && GET_MODE_2XWIDER_MODE (mode).exists (&wider_mode)
      && (convert_optab_handler ((unsignedp
				  ? umul_widen_optab
				  : smul_widen_optab),
				 wider_mode, mode) != CODE_FOR_nothing))
    {
      /* *_widen_optab needs to determine operand mode, make sure at least
	 one operand has non-VOID mode.  */
      if (GET_MODE (op0) == VOIDmode && GET_MODE (op1) == VOIDmode)
	op0 = force_reg (mode, op0);
      temp = expand_binop (wider_mode,
			   unsignedp ? umul_widen_optab : smul_widen_optab,
			   op0, op1, NULL_RTX, unsignedp, OPTAB_DIRECT);

      if (temp != 0)
	{
	  if (GET_MODE_CLASS (mode) == MODE_INT
	      && TRULY_NOOP_TRUNCATION_MODES_P (mode, GET_MODE (temp)))
	    return gen_lowpart (mode, temp);
	  else
	    return convert_to_mode (mode, temp, unsignedp);
	}
    }

  /* If this is a vector shift by a scalar, see if we can do a vector
     shift by a vector.  If so, broadcast the scalar into a vector.  */
  if (mclass == MODE_VECTOR_INT)
    {
      optab otheroptab = unknown_optab;

      if (binoptab == ashl_optab)
	otheroptab = vashl_optab;
      else if (binoptab == ashr_optab)
	otheroptab = vashr_optab;
      else if (binoptab == lshr_optab)
	otheroptab = vlshr_optab;
      else if (binoptab == rotl_optab)
	otheroptab = vrotl_optab;
      else if (binoptab == rotr_optab)
	otheroptab = vrotr_optab;

      if (otheroptab
	  && (icode = optab_handler (otheroptab, mode)) != CODE_FOR_nothing)
	{
	  /* The scalar may have been extended to be too wide.  Truncate
	     it back to the proper size to fit in the broadcast vector.  */
	  scalar_mode inner_mode = GET_MODE_INNER (mode);
	  if (!CONST_INT_P (op1)
	      && (GET_MODE_BITSIZE (as_a <scalar_int_mode> (GET_MODE (op1)))
		  > GET_MODE_BITSIZE (inner_mode)))
	    op1 = force_reg (inner_mode,
			     simplify_gen_unary (TRUNCATE, inner_mode, op1,
						 GET_MODE (op1)));
	  rtx vop1 = expand_vector_broadcast (mode, op1);
	  if (vop1)
	    {
	      temp = expand_binop_directly (icode, mode, otheroptab, op0, vop1,
					    target, unsignedp, methods, last);
	      if (temp)
		return temp;
	    }
	}
    }

  /* Look for a wider mode of the same class for which we think we
     can open-code the operation.  Check for a widening multiply at the
     wider mode as well.  */

  if (CLASS_HAS_WIDER_MODES_P (mclass)
      && methods != OPTAB_DIRECT && methods != OPTAB_LIB)
    FOR_EACH_WIDER_MODE (wider_mode, mode)
      {
	machine_mode next_mode;
	if (optab_handler (binoptab, wider_mode) != CODE_FOR_nothing
	    || (binoptab == smul_optab
		&& GET_MODE_WIDER_MODE (wider_mode).exists (&next_mode)
		&& (find_widening_optab_handler ((unsignedp
						  ? umul_widen_optab
						  : smul_widen_optab),
						 next_mode, mode)
		    != CODE_FOR_nothing)))
	  {
	    rtx xop0 = op0, xop1 = op1;
	    int no_extend = 0;

	    /* For certain integer operations, we need not actually extend
	       the narrow operands, as long as we will truncate
	       the results to the same narrowness.  */

	    if ((binoptab == ior_optab || binoptab == and_optab
		 || binoptab == xor_optab
		 || binoptab == add_optab || binoptab == sub_optab
		 || binoptab == smul_optab || binoptab == ashl_optab)
		&& mclass == MODE_INT)
	      {
		no_extend = 1;
		xop0 = avoid_expensive_constant (mode, binoptab, 0,
						 xop0, unsignedp);
		if (binoptab != ashl_optab)
		  xop1 = avoid_expensive_constant (mode, binoptab, 1,
						   xop1, unsignedp);
	      }

	    xop0 = widen_operand (xop0, wider_mode, mode, unsignedp, no_extend);

	    /* The second operand of a shift must always be extended.  */
	    xop1 = widen_operand (xop1, wider_mode, mode, unsignedp,
				  no_extend && binoptab != ashl_optab);

	    temp = expand_binop (wider_mode, binoptab, xop0, xop1, NULL_RTX,
				 unsignedp, OPTAB_DIRECT);
	    if (temp)
	      {
		if (mclass != MODE_INT
                    || !TRULY_NOOP_TRUNCATION_MODES_P (mode, wider_mode))
		  {
		    if (target == 0)
		      target = gen_reg_rtx (mode);
		    convert_move (target, temp, 0);
		    return target;
		  }
		else
		  return gen_lowpart (mode, temp);
	      }
	    else
	      delete_insns_since (last);
	  }
      }

  /* If operation is commutative,
     try to make the first operand a register.
     Even better, try to make it the same as the target.
     Also try to make the last operand a constant.  */
  if (commutative_optab_p (binoptab)
      && swap_commutative_operands_with_target (target, op0, op1))
    std::swap (op0, op1);

  /* These can be done a word at a time.  */
  if ((binoptab == and_optab || binoptab == ior_optab || binoptab == xor_optab)
      && is_int_mode (mode, &int_mode)
      && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD
      && optab_handler (binoptab, word_mode) != CODE_FOR_nothing)
    {
      int i;
      rtx_insn *insns;

      /* If TARGET is the same as one of the operands, the REG_EQUAL note
	 won't be accurate, so use a new target.  */
      if (target == 0
	  || target == op0
	  || target == op1
	  || reg_overlap_mentioned_p (target, op0)
	  || reg_overlap_mentioned_p (target, op1)
	  || !valid_multiword_target_p (target))
	target = gen_reg_rtx (int_mode);

      start_sequence ();

      /* Do the actual arithmetic.  */
      machine_mode op0_mode = GET_MODE (op0);
      machine_mode op1_mode = GET_MODE (op1);
      if (op0_mode == VOIDmode)
	op0_mode = int_mode;
      if (op1_mode == VOIDmode)
	op1_mode = int_mode;
      for (i = 0; i < GET_MODE_BITSIZE (int_mode) / BITS_PER_WORD; i++)
	{
	  rtx target_piece = operand_subword (target, i, 1, int_mode);
	  rtx x = expand_binop (word_mode, binoptab,
				operand_subword_force (op0, i, op0_mode),
				operand_subword_force (op1, i, op1_mode),
				target_piece, unsignedp, next_methods);

	  if (x == 0)
	    break;

	  if (target_piece != x)
	    emit_move_insn (target_piece, x);
	}

      insns = get_insns ();
      end_sequence ();

      if (i == GET_MODE_BITSIZE (int_mode) / BITS_PER_WORD)
	{
	  emit_insn (insns);
	  return target;
	}
    }

  /* Synthesize double word shifts from single word shifts.  */
  if ((binoptab == lshr_optab || binoptab == ashl_optab
       || binoptab == ashr_optab)
      && is_int_mode (mode, &int_mode)
      && (CONST_INT_P (op1) || optimize_insn_for_speed_p ())
      && GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD
      && GET_MODE_PRECISION (int_mode) == GET_MODE_BITSIZE (int_mode)
      && optab_handler (binoptab, word_mode) != CODE_FOR_nothing
      && optab_handler (ashl_optab, word_mode) != CODE_FOR_nothing
      && optab_handler (lshr_optab, word_mode) != CODE_FOR_nothing)
    {
      unsigned HOST_WIDE_INT shift_mask, double_shift_mask;
      scalar_int_mode op1_mode;

      double_shift_mask = targetm.shift_truncation_mask (int_mode);
      shift_mask = targetm.shift_truncation_mask (word_mode);
      op1_mode = (GET_MODE (op1) != VOIDmode
		  ? as_a <scalar_int_mode> (GET_MODE (op1))
		  : word_mode);

      /* Apply the truncation to constant shifts.  */
      if (double_shift_mask > 0 && CONST_INT_P (op1))
	op1 = gen_int_mode (INTVAL (op1) & double_shift_mask, op1_mode);

      if (op1 == CONST0_RTX (op1_mode))
	return op0;

      /* Make sure that this is a combination that expand_doubleword_shift
	 can handle.  See the comments there for details.  */
      if (double_shift_mask == 0
	  || (shift_mask == BITS_PER_WORD - 1
	      && double_shift_mask == BITS_PER_WORD * 2 - 1))
	{
	  rtx_insn *insns;
	  rtx into_target, outof_target;
	  rtx into_input, outof_input;
	  int left_shift, outof_word;

	  /* If TARGET is the same as one of the operands, the REG_EQUAL note
	     won't be accurate, so use a new target.  */
	  if (target == 0
	      || target == op0
	      || target == op1
	      || reg_overlap_mentioned_p (target, op0)
	      || reg_overlap_mentioned_p (target, op1)
	      || !valid_multiword_target_p (target))
	    target = gen_reg_rtx (int_mode);

	  start_sequence ();

	  /* OUTOF_* is the word we are shifting bits away from, and
	     INTO_* is the word that we are shifting bits towards, thus
	     they differ depending on the direction of the shift and
	     WORDS_BIG_ENDIAN.  */

	  left_shift = binoptab == ashl_optab;
	  outof_word = left_shift ^ ! WORDS_BIG_ENDIAN;

	  outof_target = operand_subword (target, outof_word, 1, int_mode);
	  into_target = operand_subword (target, 1 - outof_word, 1, int_mode);

	  outof_input = operand_subword_force (op0, outof_word, int_mode);
	  into_input = operand_subword_force (op0, 1 - outof_word, int_mode);

	  if (expand_doubleword_shift (op1_mode, binoptab,
				       outof_input, into_input, op1,
				       outof_target, into_target,
				       unsignedp, next_methods, shift_mask))
	    {
	      insns = get_insns ();
	      end_sequence ();

	      emit_insn (insns);
	      return target;
	    }
	  end_sequence ();
	}
    }

  /* Synthesize double word rotates from single word shifts.  */
  if ((binoptab == rotl_optab || binoptab == rotr_optab)
      && is_int_mode (mode, &int_mode)
      && CONST_INT_P (op1)
      && GET_MODE_PRECISION (int_mode) == 2 * BITS_PER_WORD
      && optab_handler (ashl_optab, word_mode) != CODE_FOR_nothing
      && optab_handler (lshr_optab, word_mode) != CODE_FOR_nothing)
    {
      rtx_insn *insns;
      rtx into_target, outof_target;
      rtx into_input, outof_input;
      rtx inter;
      int shift_count, left_shift, outof_word;

      /* If TARGET is the same as one of the operands, the REG_EQUAL note
	 won't be accurate, so use a new target. Do this also if target is not
	 a REG, first because having a register instead may open optimization
	 opportunities, and second because if target and op0 happen to be MEMs
	 designating the same location, we would risk clobbering it too early
	 in the code sequence we generate below.  */
      if (target == 0
	  || target == op0
	  || target == op1
	  || !REG_P (target)
	  || reg_overlap_mentioned_p (target, op0)
	  || reg_overlap_mentioned_p (target, op1)
	  || !valid_multiword_target_p (target))
	target = gen_reg_rtx (int_mode);

      start_sequence ();

      shift_count = INTVAL (op1);

      /* OUTOF_* is the word we are shifting bits away from, and
	 INTO_* is the word that we are shifting bits towards, thus
	 they differ depending on the direction of the shift and
	 WORDS_BIG_ENDIAN.  */

      left_shift = (binoptab == rotl_optab);
      outof_word = left_shift ^ ! WORDS_BIG_ENDIAN;

      outof_target = operand_subword (target, outof_word, 1, int_mode);
      into_target = operand_subword (target, 1 - outof_word, 1, int_mode);

      outof_input = operand_subword_force (op0, outof_word, int_mode);
      into_input = operand_subword_force (op0, 1 - outof_word, int_mode);

      if (shift_count == BITS_PER_WORD)
	{
	  /* This is just a word swap.  */
	  emit_move_insn (outof_target, into_input);
	  emit_move_insn (into_target, outof_input);
	  inter = const0_rtx;
	}
      else
	{
	  rtx into_temp1, into_temp2, outof_temp1, outof_temp2;
	  HOST_WIDE_INT first_shift_count, second_shift_count;
	  optab reverse_unsigned_shift, unsigned_shift;

	  reverse_unsigned_shift = (left_shift ^ (shift_count < BITS_PER_WORD)
				    ? lshr_optab : ashl_optab);

	  unsigned_shift = (left_shift ^ (shift_count < BITS_PER_WORD)
			    ? ashl_optab : lshr_optab);

	  if (shift_count > BITS_PER_WORD)
	    {
	      first_shift_count = shift_count - BITS_PER_WORD;
	      second_shift_count = 2 * BITS_PER_WORD - shift_count;
	    }
	  else
	    {
	      first_shift_count = BITS_PER_WORD - shift_count;
	      second_shift_count = shift_count;
	    }
	  rtx first_shift_count_rtx
	    = gen_int_shift_amount (word_mode, first_shift_count);
	  rtx second_shift_count_rtx
	    = gen_int_shift_amount (word_mode, second_shift_count);

	  into_temp1 = expand_binop (word_mode, unsigned_shift,
				     outof_input, first_shift_count_rtx,
				     NULL_RTX, unsignedp, next_methods);
	  into_temp2 = expand_binop (word_mode, reverse_unsigned_shift,
				     into_input, second_shift_count_rtx,
				     NULL_RTX, unsignedp, next_methods);

	  if (into_temp1 != 0 && into_temp2 != 0)
	    inter = expand_binop (word_mode, ior_optab, into_temp1, into_temp2,
				  into_target, unsignedp, next_methods);
	  else
	    inter = 0;

	  if (inter != 0 && inter != into_target)
	    emit_move_insn (into_target, inter);

	  outof_temp1 = expand_binop (word_mode, unsigned_shift,
				      into_input, first_shift_count_rtx,
				      NULL_RTX, unsignedp, next_methods);
	  outof_temp2 = expand_binop (word_mode, reverse_unsigned_shift,
				      outof_input, second_shift_count_rtx,
				      NULL_RTX, unsignedp, next_methods);

	  if (inter != 0 && outof_temp1 != 0 && outof_temp2 != 0)
	    inter = expand_binop (word_mode, ior_optab,
				  outof_temp1, outof_temp2,
				  outof_target, unsignedp, next_methods);

	  if (inter != 0 && inter != outof_target)
	    emit_move_insn (outof_target, inter);
	}

      insns = get_insns ();
      end_sequence ();

      if (inter != 0)
	{
	  emit_insn (insns);
	  return target;
	}
    }

  /* These can be done a word at a time by propagating carries.  */
  if ((binoptab == add_optab || binoptab == sub_optab)
      && is_int_mode (mode, &int_mode)
      && GET_MODE_SIZE (int_mode) >= 2 * UNITS_PER_WORD
      && optab_handler (binoptab, word_mode) != CODE_FOR_nothing)
    {
      unsigned int i;
      optab otheroptab = binoptab == add_optab ? sub_optab : add_optab;
      const unsigned int nwords = GET_MODE_BITSIZE (int_mode) / BITS_PER_WORD;
      rtx carry_in = NULL_RTX, carry_out = NULL_RTX;
      rtx xop0, xop1, xtarget;

      /* We can handle either a 1 or -1 value for the carry.  If STORE_FLAG
	 value is one of those, use it.  Otherwise, use 1 since it is the
	 one easiest to get.  */
#if STORE_FLAG_VALUE == 1 || STORE_FLAG_VALUE == -1
      int normalizep = STORE_FLAG_VALUE;
#else
      int normalizep = 1;
#endif

      /* Prepare the operands.  */
      xop0 = force_reg (int_mode, op0);
      xop1 = force_reg (int_mode, op1);

      xtarget = gen_reg_rtx (int_mode);

      if (target == 0 || !REG_P (target) || !valid_multiword_target_p (target))
	target = xtarget;

      /* Indicate for flow that the entire target reg is being set.  */
      if (REG_P (target))
	emit_clobber (xtarget);

      /* Do the actual arithmetic.  */
      for (i = 0; i < nwords; i++)
	{
	  int index = (WORDS_BIG_ENDIAN ? nwords - i - 1 : i);
	  rtx target_piece = operand_subword (xtarget, index, 1, int_mode);
	  rtx op0_piece = operand_subword_force (xop0, index, int_mode);
	  rtx op1_piece = operand_subword_force (xop1, index, int_mode);
	  rtx x;

	  /* Main add/subtract of the input operands.  */
	  x = expand_binop (word_mode, binoptab,
			    op0_piece, op1_piece,
			    target_piece, unsignedp, next_methods);
	  if (x == 0)
	    break;

	  if (i + 1 < nwords)
	    {
	      /* Store carry from main add/subtract.  */
	      carry_out = gen_reg_rtx (word_mode);
	      carry_out = emit_store_flag_force (carry_out,
						 (binoptab == add_optab
						  ? LT : GT),
						 x, op0_piece,
						 word_mode, 1, normalizep);
	    }

	  if (i > 0)
	    {
	      rtx newx;

	      /* Add/subtract previous carry to main result.  */
	      newx = expand_binop (word_mode,
				   normalizep == 1 ? binoptab : otheroptab,
				   x, carry_in,
				   NULL_RTX, 1, next_methods);

	      if (i + 1 < nwords)
		{
		  /* Get out carry from adding/subtracting carry in.  */
		  rtx carry_tmp = gen_reg_rtx (word_mode);
		  carry_tmp = emit_store_flag_force (carry_tmp,
						     (binoptab == add_optab
						      ? LT : GT),
						     newx, x,
						     word_mode, 1, normalizep);

		  /* Logical-ior the two poss. carry together.  */
		  carry_out = expand_binop (word_mode, ior_optab,
					    carry_out, carry_tmp,
					    carry_out, 0, next_methods);
		  if (carry_out == 0)
		    break;
		}
	      emit_move_insn (target_piece, newx);
	    }
	  else
	    {
	      if (x != target_piece)
		emit_move_insn (target_piece, x);
	    }

	  carry_in = carry_out;
	}

      if (i == GET_MODE_BITSIZE (int_mode) / (unsigned) BITS_PER_WORD)
	{
	  if (optab_handler (mov_optab, int_mode) != CODE_FOR_nothing
	      || ! rtx_equal_p (target, xtarget))
	    {
	      rtx_insn *temp = emit_move_insn (target, xtarget);

	      set_dst_reg_note (temp, REG_EQUAL,
				gen_rtx_fmt_ee (optab_to_code (binoptab),
						int_mode, copy_rtx (xop0),
						copy_rtx (xop1)),
				target);
	    }
	  else
	    target = xtarget;

	  return target;
	}

      else
	delete_insns_since (last);
    }

  /* Attempt to synthesize double word multiplies using a sequence of word
     mode multiplications.  We first attempt to generate a sequence using a
     more efficient unsigned widening multiply, and if that fails we then
     try using a signed widening multiply.  */

  if (binoptab == smul_optab
      && is_int_mode (mode, &int_mode)
      && GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD
      && optab_handler (smul_optab, word_mode) != CODE_FOR_nothing
      && optab_handler (add_optab, word_mode) != CODE_FOR_nothing)
    {
      rtx product = NULL_RTX;
      if (convert_optab_handler (umul_widen_optab, int_mode, word_mode)
	  != CODE_FOR_nothing)
	{
	  product = expand_doubleword_mult (int_mode, op0, op1, target,
					    true, methods);
	  if (!product)
	    delete_insns_since (last);
	}

      if (product == NULL_RTX
	  && (convert_optab_handler (smul_widen_optab, int_mode, word_mode)
	      != CODE_FOR_nothing))
	{
	  product = expand_doubleword_mult (int_mode, op0, op1, target,
					    false, methods);
	  if (!product)
	    delete_insns_since (last);
	}

      if (product != NULL_RTX)
	{
	  if (optab_handler (mov_optab, int_mode) != CODE_FOR_nothing)
	    {
	      rtx_insn *move = emit_move_insn (target ? target : product,
					       product);
	      set_dst_reg_note (move,
				REG_EQUAL,
				gen_rtx_fmt_ee (MULT, int_mode,
						copy_rtx (op0),
						copy_rtx (op1)),
				target ? target : product);
	    }
	  return product;
	}
    }

  /* It can't be open-coded in this mode.
     Use a library call if one is available and caller says that's ok.  */

  libfunc = optab_libfunc (binoptab, mode);
  if (libfunc
      && (methods == OPTAB_LIB || methods == OPTAB_LIB_WIDEN))
    {
      rtx_insn *insns;
      rtx op1x = op1;
      machine_mode op1_mode = mode;
      rtx value;

      start_sequence ();

      if (shift_optab_p (binoptab))
	{
	  op1_mode = targetm.libgcc_shift_count_mode ();
	  /* Specify unsigned here,
	     since negative shift counts are meaningless.  */
	  op1x = convert_to_mode (op1_mode, op1, 1);
	}

      if (GET_MODE (op0) != VOIDmode
	  && GET_MODE (op0) != mode)
	op0 = convert_to_mode (mode, op0, unsignedp);

      /* Pass 1 for NO_QUEUE so we don't lose any increments
	 if the libcall is cse'd or moved.  */
      value = emit_library_call_value (libfunc,
				       NULL_RTX, LCT_CONST, mode,
				       op0, mode, op1x, op1_mode);

      insns = get_insns ();
      end_sequence ();

      bool trapv = trapv_binoptab_p (binoptab);
      target = gen_reg_rtx (mode);
      emit_libcall_block_1 (insns, target, value,
			    trapv ? NULL_RTX
			    : gen_rtx_fmt_ee (optab_to_code (binoptab),
					      mode, op0, op1), trapv);

      return target;
    }

  delete_insns_since (last);

  /* It can't be done in this mode.  Can we do it in a wider mode?  */

  if (! (methods == OPTAB_WIDEN || methods == OPTAB_LIB_WIDEN
	 || methods == OPTAB_MUST_WIDEN))
    {
      /* Caller says, don't even try.  */
      delete_insns_since (entry_last);
      return 0;
    }

  /* Compute the value of METHODS to pass to recursive calls.
     Don't allow widening to be tried recursively.  */

  methods = (methods == OPTAB_LIB_WIDEN ? OPTAB_LIB : OPTAB_DIRECT);

  /* Look for a wider mode of the same class for which it appears we can do
     the operation.  */

  if (CLASS_HAS_WIDER_MODES_P (mclass))
    {
      /* This code doesn't make sense for conversion optabs, since we
	 wouldn't then want to extend the operands to be the same size
	 as the result.  */
      gcc_assert (!convert_optab_p (binoptab));
      FOR_EACH_WIDER_MODE (wider_mode, mode)
	{
	  if (optab_handler (binoptab, wider_mode)
	      || (methods == OPTAB_LIB
		  && optab_libfunc (binoptab, wider_mode)))
	    {
	      rtx xop0 = op0, xop1 = op1;
	      int no_extend = 0;

	      /* For certain integer operations, we need not actually extend
		 the narrow operands, as long as we will truncate
		 the results to the same narrowness.  */

	      if ((binoptab == ior_optab || binoptab == and_optab
		   || binoptab == xor_optab
		   || binoptab == add_optab || binoptab == sub_optab
		   || binoptab == smul_optab || binoptab == ashl_optab)
		  && mclass == MODE_INT)
		no_extend = 1;

	      xop0 = widen_operand (xop0, wider_mode, mode,
				    unsignedp, no_extend);

	      /* The second operand of a shift must always be extended.  */
	      xop1 = widen_operand (xop1, wider_mode, mode, unsignedp,
				    no_extend && binoptab != ashl_optab);

	      temp = expand_binop (wider_mode, binoptab, xop0, xop1, NULL_RTX,
				   unsignedp, methods);
	      if (temp)
		{
		  if (mclass != MODE_INT
		      || !TRULY_NOOP_TRUNCATION_MODES_P (mode, wider_mode))
		    {
		      if (target == 0)
			target = gen_reg_rtx (mode);
		      convert_move (target, temp, 0);
		      return target;
		    }
		  else
		    return gen_lowpart (mode, temp);
		}
	      else
		delete_insns_since (last);
	    }
	}
    }

  delete_insns_since (entry_last);
  return 0;
}

/* Expand a binary operator which has both signed and unsigned forms.
   UOPTAB is the optab for unsigned operations, and SOPTAB is for
   signed operations.

   If we widen unsigned operands, we may use a signed wider operation instead
   of an unsigned wider operation, since the result would be the same.  */

rtx
sign_expand_binop (machine_mode mode, optab uoptab, optab soptab,
		   rtx op0, rtx op1, rtx target, int unsignedp,
		   enum optab_methods methods)
{
  rtx temp;
  optab direct_optab = unsignedp ? uoptab : soptab;
  bool save_enable;

  /* Do it without widening, if possible.  */
  temp = expand_binop (mode, direct_optab, op0, op1, target,
		       unsignedp, OPTAB_DIRECT);
  if (temp || methods == OPTAB_DIRECT)
    return temp;

  /* Try widening to a signed int.  Disable any direct use of any
     signed insn in the current mode.  */
  save_enable = swap_optab_enable (soptab, mode, false);

  temp = expand_binop (mode, soptab, op0, op1, target,
		       unsignedp, OPTAB_WIDEN);

  /* For unsigned operands, try widening to an unsigned int.  */
  if (!temp && unsignedp)
    temp = expand_binop (mode, uoptab, op0, op1, target,
			 unsignedp, OPTAB_WIDEN);
  if (temp || methods == OPTAB_WIDEN)
    goto egress;

  /* Use the right width libcall if that exists.  */
  temp = expand_binop (mode, direct_optab, op0, op1, target,
		       unsignedp, OPTAB_LIB);
  if (temp || methods == OPTAB_LIB)
    goto egress;

  /* Must widen and use a libcall, use either signed or unsigned.  */
  temp = expand_binop (mode, soptab, op0, op1, target,
		       unsignedp, methods);
  if (!temp && unsignedp)
    temp = expand_binop (mode, uoptab, op0, op1, target,
			 unsignedp, methods);

 egress:
  /* Undo the fiddling above.  */
  if (save_enable)
    swap_optab_enable (soptab, mode, true);
  return temp;
}

/* Generate code to perform an operation specified by UNOPPTAB
   on operand OP0, with two results to TARG0 and TARG1.
   We assume that the order of the operands for the instruction
   is TARG0, TARG1, OP0.

   Either TARG0 or TARG1 may be zero, but what that means is that
   the result is not actually wanted.  We will generate it into
   a dummy pseudo-reg and discard it.  They may not both be zero.

   Returns 1 if this operation can be performed; 0 if not.  */

int
expand_twoval_unop (optab unoptab, rtx op0, rtx targ0, rtx targ1,
		    int unsignedp)
{
  machine_mode mode = GET_MODE (targ0 ? targ0 : targ1);
  enum mode_class mclass;
  machine_mode wider_mode;
  rtx_insn *entry_last = get_last_insn ();
  rtx_insn *last;

  mclass = GET_MODE_CLASS (mode);

  if (!targ0)
    targ0 = gen_reg_rtx (mode);
  if (!targ1)
    targ1 = gen_reg_rtx (mode);

  /* Record where to go back to if we fail.  */
  last = get_last_insn ();

  if (optab_handler (unoptab, mode) != CODE_FOR_nothing)
    {
      struct expand_operand ops[3];
      enum insn_code icode = optab_handler (unoptab, mode);

      create_fixed_operand (&ops[0], targ0);
      create_fixed_operand (&ops[1], targ1);
      create_convert_operand_from (&ops[2], op0, mode, unsignedp);
      if (maybe_expand_insn (icode, 3, ops))
	return 1;
    }

  /* It can't be done in this mode.  Can we do it in a wider mode?  */

  if (CLASS_HAS_WIDER_MODES_P (mclass))
    {
      FOR_EACH_WIDER_MODE (wider_mode, mode)
	{
	  if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing)
	    {
	      rtx t0 = gen_reg_rtx (wider_mode);
	      rtx t1 = gen_reg_rtx (wider_mode);
	      rtx cop0 = convert_modes (wider_mode, mode, op0, unsignedp);

	      if (expand_twoval_unop (unoptab, cop0, t0, t1, unsignedp))
		{
		  convert_move (targ0, t0, unsignedp);
		  convert_move (targ1, t1, unsignedp);
		  return 1;
		}
	      else
		delete_insns_since (last);
	    }
	}
    }

  delete_insns_since (entry_last);
  return 0;
}

/* Generate code to perform an operation specified by BINOPTAB
   on operands OP0 and OP1, with two results to TARG1 and TARG2.
   We assume that the order of the operands for the instruction
   is TARG0, OP0, OP1, TARG1, which would fit a pattern like
   [(set TARG0 (operate OP0 OP1)) (set TARG1 (operate ...))].

   Either TARG0 or TARG1 may be zero, but what that means is that
   the result is not actually wanted.  We will generate it into
   a dummy pseudo-reg and discard it.  They may not both be zero.

   Returns 1 if this operation can be performed; 0 if not.  */

int
expand_twoval_binop (optab binoptab, rtx op0, rtx op1, rtx targ0, rtx targ1,
		     int unsignedp)
{
  machine_mode mode = GET_MODE (targ0 ? targ0 : targ1);
  enum mode_class mclass;
  machine_mode wider_mode;
  rtx_insn *entry_last = get_last_insn ();
  rtx_insn *last;

  mclass = GET_MODE_CLASS (mode);

  if (!targ0)
    targ0 = gen_reg_rtx (mode);
  if (!targ1)
    targ1 = gen_reg_rtx (mode);

  /* Record where to go back to if we fail.  */
  last = get_last_insn ();

  if (optab_handler (binoptab, mode) != CODE_FOR_nothing)
    {
      struct expand_operand ops[4];
      enum insn_code icode = optab_handler (binoptab, mode);
      machine_mode mode0 = insn_data[icode].operand[1].mode;
      machine_mode mode1 = insn_data[icode].operand[2].mode;
      rtx xop0 = op0, xop1 = op1;

      /* If we are optimizing, force expensive constants into a register.  */
      xop0 = avoid_expensive_constant (mode0, binoptab, 0, xop0, unsignedp);
      xop1 = avoid_expensive_constant (mode1, binoptab, 1, xop1, unsignedp);

      create_fixed_operand (&ops[0], targ0);
      create_convert_operand_from (&ops[1], op0, mode, unsignedp);
      create_convert_operand_from (&ops[2], op1, mode, unsignedp);
      create_fixed_operand (&ops[3], targ1);
      if (maybe_expand_insn (icode, 4, ops))
	return 1;
      delete_insns_since (last);
    }

  /* It can't be done in this mode.  Can we do it in a wider mode?  */

  if (CLASS_HAS_WIDER_MODES_P (mclass))
    {
      FOR_EACH_WIDER_MODE (wider_mode, mode)
	{
	  if (optab_handler (binoptab, wider_mode) != CODE_FOR_nothing)
	    {
	      rtx t0 = gen_reg_rtx (wider_mode);
	      rtx t1 = gen_reg_rtx (wider_mode);
	      rtx cop0 = convert_modes (wider_mode, mode, op0, unsignedp);
	      rtx cop1 = convert_modes (wider_mode, mode, op1, unsignedp);

	      if (expand_twoval_binop (binoptab, cop0, cop1,
				       t0, t1, unsignedp))
		{
		  convert_move (targ0, t0, unsignedp);
		  convert_move (targ1, t1, unsignedp);
		  return 1;
		}
	      else
		delete_insns_since (last);
	    }
	}
    }

  delete_insns_since (entry_last);
  return 0;
}

/* Expand the two-valued library call indicated by BINOPTAB, but
   preserve only one of the values.  If TARG0 is non-NULL, the first
   value is placed into TARG0; otherwise the second value is placed
   into TARG1.  Exactly one of TARG0 and TARG1 must be non-NULL.  The
   value stored into TARG0 or TARG1 is equivalent to (CODE OP0 OP1).
   This routine assumes that the value returned by the library call is
   as if the return value was of an integral mode twice as wide as the
   mode of OP0.  Returns 1 if the call was successful.  */

bool
expand_twoval_binop_libfunc (optab binoptab, rtx op0, rtx op1,
			     rtx targ0, rtx targ1, enum rtx_code code)
{
  machine_mode mode;
  machine_mode libval_mode;
  rtx libval;
  rtx_insn *insns;
  rtx libfunc;

  /* Exactly one of TARG0 or TARG1 should be non-NULL.  */
  gcc_assert (!targ0 != !targ1);

  mode = GET_MODE (op0);
  libfunc = optab_libfunc (binoptab, mode);
  if (!libfunc)
    return false;

  /* The value returned by the library function will have twice as
     many bits as the nominal MODE.  */
  libval_mode = smallest_int_mode_for_size (2 * GET_MODE_BITSIZE (mode));
  start_sequence ();
  libval = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST,
				    libval_mode,
				    op0, mode,
				    op1, mode);
  /* Get the part of VAL containing the value that we want.  */
  libval = simplify_gen_subreg (mode, libval, libval_mode,
				targ0 ? 0 : GET_MODE_SIZE (mode));
  insns = get_insns ();
  end_sequence ();
  /* Move the into the desired location.  */
  emit_libcall_block (insns, targ0 ? targ0 : targ1, libval,
		      gen_rtx_fmt_ee (code, mode, op0, op1));

  return true;
}


/* Wrapper around expand_unop which takes an rtx code to specify
   the operation to perform, not an optab pointer.  All other
   arguments are the same.  */
rtx
expand_simple_unop (machine_mode mode, enum rtx_code code, rtx op0,
		    rtx target, int unsignedp)
{
  optab unop = code_to_optab (code);
  gcc_assert (unop);

  return expand_unop (mode, unop, op0, target, unsignedp);
}

/* Try calculating
	(clz:narrow x)
   as
	(clz:wide (zero_extend:wide x)) - ((width wide) - (width narrow)).

   A similar operation can be used for clrsb.  UNOPTAB says which operation
   we are trying to expand.  */
static rtx
widen_leading (scalar_int_mode mode, rtx op0, rtx target, optab unoptab)
{
  opt_scalar_int_mode wider_mode_iter;
  FOR_EACH_WIDER_MODE (wider_mode_iter, mode)
    {
      scalar_int_mode wider_mode = wider_mode_iter.require ();
      if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing)
	{
	  rtx xop0, temp;
	  rtx_insn *last;

	  last = get_last_insn ();

	  if (target == 0)
	    target = gen_reg_rtx (mode);
	  xop0 = widen_operand (op0, wider_mode, mode,
				unoptab != clrsb_optab, false);
	  temp = expand_unop (wider_mode, unoptab, xop0, NULL_RTX,
			      unoptab != clrsb_optab);
	  if (temp != 0)
	    temp = expand_binop
	      (wider_mode, sub_optab, temp,
	       gen_int_mode (GET_MODE_PRECISION (wider_mode)
			     - GET_MODE_PRECISION (mode),
			     wider_mode),
	       target, true, OPTAB_DIRECT);
	  if (temp == 0)
	    delete_insns_since (last);

	  return temp;
	}
    }
  return 0;
}

/* Try calculating clz of a double-word quantity as two clz's of word-sized
   quantities, choosing which based on whether the high word is nonzero.  */
static rtx
expand_doubleword_clz (scalar_int_mode mode, rtx op0, rtx target)
{
  rtx xop0 = force_reg (mode, op0);
  rtx subhi = gen_highpart (word_mode, xop0);
  rtx sublo = gen_lowpart (word_mode, xop0);
  rtx_code_label *hi0_label = gen_label_rtx ();
  rtx_code_label *after_label = gen_label_rtx ();
  rtx_insn *seq;
  rtx temp, result;

  /* If we were not given a target, use a word_mode register, not a
     'mode' register.  The result will fit, and nobody is expecting
     anything bigger (the return type of __builtin_clz* is int).  */
  if (!target)
    target = gen_reg_rtx (word_mode);

  /* In any case, write to a word_mode scratch in both branches of the
     conditional, so we can ensure there is a single move insn setting
     'target' to tag a REG_EQUAL note on.  */
  result = gen_reg_rtx (word_mode);

  start_sequence ();

  /* If the high word is not equal to zero,
     then clz of the full value is clz of the high word.  */
  emit_cmp_and_jump_insns (subhi, CONST0_RTX (word_mode), EQ, 0,
			   word_mode, true, hi0_label);

  temp = expand_unop_direct (word_mode, clz_optab, subhi, result, true);
  if (!temp)
    goto fail;

  if (temp != result)
    convert_move (result, temp, true);

  emit_jump_insn (targetm.gen_jump (after_label));
  emit_barrier ();

  /* Else clz of the full value is clz of the low word plus the number
     of bits in the high word.  */
  emit_label (hi0_label);

  temp = expand_unop_direct (word_mode, clz_optab, sublo, 0, true);
  if (!temp)
    goto fail;
  temp = expand_binop (word_mode, add_optab, temp,
		       gen_int_mode (GET_MODE_BITSIZE (word_mode), word_mode),
		       result, true, OPTAB_DIRECT);
  if (!temp)
    goto fail;
  if (temp != result)
    convert_move (result, temp, true);

  emit_label (after_label);
  convert_move (target, result, true);

  seq = get_insns ();
  end_sequence ();

  add_equal_note (seq, target, CLZ, xop0, NULL_RTX, mode);
  emit_insn (seq);
  return target;

 fail:
  end_sequence ();
  return 0;
}

/* Try calculating popcount of a double-word quantity as two popcount's of
   word-sized quantities and summing up the results.  */
static rtx
expand_doubleword_popcount (scalar_int_mode mode, rtx op0, rtx target)
{
  rtx t0, t1, t;
  rtx_insn *seq;

  start_sequence ();

  t0 = expand_unop_direct (word_mode, popcount_optab,
			   operand_subword_force (op0, 0, mode), NULL_RTX,
			   true);
  t1 = expand_unop_direct (word_mode, popcount_optab,
			   operand_subword_force (op0, 1, mode), NULL_RTX,
			   true);
  if (!t0 || !t1)
    {
      end_sequence ();
      return NULL_RTX;
    }

  /* If we were not given a target, use a word_mode register, not a
     'mode' register.  The result will fit, and nobody is expecting
     anything bigger (the return type of __builtin_popcount* is int).  */
  if (!target)
    target = gen_reg_rtx (word_mode);

  t = expand_binop (word_mode, add_optab, t0, t1, target, 0, OPTAB_DIRECT);

  seq = get_insns ();
  end_sequence ();

  add_equal_note (seq, t, POPCOUNT, op0, NULL_RTX, mode);
  emit_insn (seq);
  return t;
}

/* Try calculating
	(parity:wide x)
   as
	(parity:narrow (low (x) ^ high (x))) */
static rtx
expand_doubleword_parity (scalar_int_mode mode, rtx op0, rtx target)
{
  rtx t = expand_binop (word_mode, xor_optab,
			operand_subword_force (op0, 0, mode),
			operand_subword_force (op0, 1, mode),
			NULL_RTX, 0, OPTAB_DIRECT);
  return expand_unop (word_mode, parity_optab, t, target, true);
}

/* Try calculating
	(bswap:narrow x)
   as
	(lshiftrt:wide (bswap:wide x) ((width wide) - (width narrow))).  */
static rtx
widen_bswap (scalar_int_mode mode, rtx op0, rtx target)
{
  rtx x;
  rtx_insn *last;
  opt_scalar_int_mode wider_mode_iter;

  FOR_EACH_WIDER_MODE (wider_mode_iter, mode)
    if (optab_handler (bswap_optab, wider_mode_iter.require ())
	!= CODE_FOR_nothing)
      break;

  if (!wider_mode_iter.exists ())
    return NULL_RTX;

  scalar_int_mode wider_mode = wider_mode_iter.require ();
  last = get_last_insn ();

  x = widen_operand (op0, wider_mode, mode, true, true);
  x = expand_unop (wider_mode, bswap_optab, x, NULL_RTX, true);

  gcc_assert (GET_MODE_PRECISION (wider_mode) == GET_MODE_BITSIZE (wider_mode)
	      && GET_MODE_PRECISION (mode) == GET_MODE_BITSIZE (mode));
  if (x != 0)
    x = expand_shift (RSHIFT_EXPR, wider_mode, x,
		      GET_MODE_BITSIZE (wider_mode)
		      - GET_MODE_BITSIZE (mode),
		      NULL_RTX, true);

  if (x != 0)
    {
      if (target == 0)
	target = gen_reg_rtx (mode);
      emit_move_insn (target, gen_lowpart (mode, x));
    }
  else
    delete_insns_since (last);

  return target;
}

/* Try calculating bswap as two bswaps of two word-sized operands.  */

static rtx
expand_doubleword_bswap (machine_mode mode, rtx op, rtx target)
{
  rtx t0, t1;

  t1 = expand_unop (word_mode, bswap_optab,
		    operand_subword_force (op, 0, mode), NULL_RTX, true);
  t0 = expand_unop (word_mode, bswap_optab,
		    operand_subword_force (op, 1, mode), NULL_RTX, true);

  if (target == 0 || !valid_multiword_target_p (target))
    target = gen_reg_rtx (mode);
  if (REG_P (target))
    emit_clobber (target);
  emit_move_insn (operand_subword (target, 0, 1, mode), t0);
  emit_move_insn (operand_subword (target, 1, 1, mode), t1);

  return target;
}

/* Try calculating (parity x) as (and (popcount x) 1), where
   popcount can also be done in a wider mode.  */
static rtx
expand_parity (scalar_int_mode mode, rtx op0, rtx target)
{
  enum mode_class mclass = GET_MODE_CLASS (mode);
  opt_scalar_int_mode wider_mode_iter;
  FOR_EACH_MODE_FROM (wider_mode_iter, mode)
    {
      scalar_int_mode wider_mode = wider_mode_iter.require ();
      if (optab_handler (popcount_optab, wider_mode) != CODE_FOR_nothing)
	{
	  rtx xop0, temp;
	  rtx_insn *last;

	  last = get_last_insn ();

	  if (target == 0 || GET_MODE (target) != wider_mode)
	    target = gen_reg_rtx (wider_mode);

	  xop0 = widen_operand (op0, wider_mode, mode, true, false);
	  temp = expand_unop (wider_mode, popcount_optab, xop0, NULL_RTX,
			      true);
	  if (temp != 0)
	    temp = expand_binop (wider_mode, and_optab, temp, const1_rtx,
				 target, true, OPTAB_DIRECT);

	  if (temp)
	    {
	      if (mclass != MODE_INT
		  || !TRULY_NOOP_TRUNCATION_MODES_P (mode, wider_mode))
		return convert_to_mode (mode, temp, 0);
	      else
		return gen_lowpart (mode, temp);
	    }
	  else
	    delete_insns_since (last);
	}
    }
  return 0;
}

/* Try calculating ctz(x) as K - clz(x & -x) ,
   where K is GET_MODE_PRECISION(mode) - 1.

   Both __builtin_ctz and __builtin_clz are undefined at zero, so we
   don't have to worry about what the hardware does in that case.  (If
   the clz instruction produces the usual value at 0, which is K, the
   result of this code sequence will be -1; expand_ffs, below, relies
   on this.  It might be nice to have it be K instead, for consistency
   with the (very few) processors that provide a ctz with a defined
   value, but that would take one more instruction, and it would be
   less convenient for expand_ffs anyway.  */

static rtx
expand_ctz (scalar_int_mode mode, rtx op0, rtx target)
{
  rtx_insn *seq;
  rtx temp;

  if (optab_handler (clz_optab, mode) == CODE_FOR_nothing)
    return 0;

  start_sequence ();

  temp = expand_unop_direct (mode, neg_optab, op0, NULL_RTX, true);
  if (temp)
    temp = expand_binop (mode, and_optab, op0, temp, NULL_RTX,
			 true, OPTAB_DIRECT);
  if (temp)
    temp = expand_unop_direct (mode, clz_optab, temp, NULL_RTX, true);
  if (temp)
    temp = expand_binop (mode, sub_optab,
			 gen_int_mode (GET_MODE_PRECISION (mode) - 1, mode),
			 temp, target,
			 true, OPTAB_DIRECT);
  if (temp == 0)
    {
      end_sequence ();
      return 0;
    }

  seq = get_insns ();
  end_sequence ();

  add_equal_note (seq, temp, CTZ, op0, NULL_RTX, mode);
  emit_insn (seq);
  return temp;
}


/* Try calculating ffs(x) using ctz(x) if we have that instruction, or
   else with the sequence used by expand_clz.

   The ffs builtin promises to return zero for a zero value and ctz/clz
   may have an undefined value in that case.  If they do not give us a
   convenient value, we have to generate a test and branch.  */
static rtx
expand_ffs (scalar_int_mode mode, rtx op0, rtx target)
{
  HOST_WIDE_INT val = 0;
  bool defined_at_zero = false;
  rtx temp;
  rtx_insn *seq;

  if (optab_handler (ctz_optab, mode) != CODE_FOR_nothing)
    {
      start_sequence ();

      temp = expand_unop_direct (mode, ctz_optab, op0, 0, true);
      if (!temp)
	goto fail;

      defined_at_zero = (CTZ_DEFINED_VALUE_AT_ZERO (mode, val) == 2);
    }
  else if (optab_handler (clz_optab, mode) != CODE_FOR_nothing)
    {
      start_sequence ();
      temp = expand_ctz (mode, op0, 0);
      if (!temp)
	goto fail;

      if (CLZ_DEFINED_VALUE_AT_ZERO (mode, val) == 2)
	{
	  defined_at_zero = true;
	  val = (GET_MODE_PRECISION (mode) - 1) - val;
	}
    }
  else
    return 0;

  if (defined_at_zero && val == -1)
    /* No correction needed at zero.  */;
  else
    {
      /* We don't try to do anything clever with the situation found
	 on some processors (eg Alpha) where ctz(0:mode) ==
	 bitsize(mode).  If someone can think of a way to send N to -1
	 and leave alone all values in the range 0..N-1 (where N is a
	 power of two), cheaper than this test-and-branch, please add it.

	 The test-and-branch is done after the operation itself, in case
	 the operation sets condition codes that can be recycled for this.
	 (This is true on i386, for instance.)  */

      rtx_code_label *nonzero_label = gen_label_rtx ();
      emit_cmp_and_jump_insns (op0, CONST0_RTX (mode), NE, 0,
			       mode, true, nonzero_label);

      convert_move (temp, GEN_INT (-1), false);
      emit_label (nonzero_label);
    }

  /* temp now has a value in the range -1..bitsize-1.  ffs is supposed
     to produce a value in the range 0..bitsize.  */
  temp = expand_binop (mode, add_optab, temp, gen_int_mode (1, mode),
		       target, false, OPTAB_DIRECT);
  if (!temp)
    goto fail;

  seq = get_insns ();
  end_sequence ();

  add_equal_note (seq, temp, FFS, op0, NULL_RTX, mode);
  emit_insn (seq);
  return temp;

 fail:
  end_sequence ();
  return 0;
}

/* Extract the OMODE lowpart from VAL, which has IMODE.  Under certain
   conditions, VAL may already be a SUBREG against which we cannot generate
   a further SUBREG.  In this case, we expect forcing the value into a
   register will work around the situation.  */

static rtx
lowpart_subreg_maybe_copy (machine_mode omode, rtx val,
			   machine_mode imode)
{
  rtx ret;
  ret = lowpart_subreg (omode, val, imode);
  if (ret == NULL)
    {
      val = force_reg (imode, val);
      ret = lowpart_subreg (omode, val, imode);
      gcc_assert (ret != NULL);
    }
  return ret;
}

/* Expand a floating point absolute value or negation operation via a
   logical operation on the sign bit.  */

static rtx
expand_absneg_bit (enum rtx_code code, scalar_float_mode mode,
		   rtx op0, rtx target)
{
  const struct real_format *fmt;
  int bitpos, word, nwords, i;
  scalar_int_mode imode;
  rtx temp;
  rtx_insn *insns;

  /* The format has to have a simple sign bit.  */
  fmt = REAL_MODE_FORMAT (mode);
  if (fmt == NULL)
    return NULL_RTX;

  bitpos = fmt->signbit_rw;
  if (bitpos < 0)
    return NULL_RTX;

  /* Don't create negative zeros if the format doesn't support them.  */
  if (code == NEG && !fmt->has_signed_zero)
    return NULL_RTX;

  if (GET_MODE_SIZE (mode) <= UNITS_PER_WORD)
    {
      if (!int_mode_for_mode (mode).exists (&imode))
	return NULL_RTX;
      word = 0;
      nwords = 1;
    }
  else
    {
      imode = word_mode;

      if (FLOAT_WORDS_BIG_ENDIAN)
	word = (GET_MODE_BITSIZE (mode) - bitpos) / BITS_PER_WORD;
      else
	word = bitpos / BITS_PER_WORD;
      bitpos = bitpos % BITS_PER_WORD;
      nwords = (GET_MODE_BITSIZE (mode) + BITS_PER_WORD - 1) / BITS_PER_WORD;
    }

  wide_int mask = wi::set_bit_in_zero (bitpos, GET_MODE_PRECISION (imode));
  if (code == ABS)
    mask = ~mask;

  if (target == 0
      || target == op0
      || reg_overlap_mentioned_p (target, op0)
      || (nwords > 1 && !valid_multiword_target_p (target)))
    target = gen_reg_rtx (mode);

  if (nwords > 1)
    {
      start_sequence ();

      for (i = 0; i < nwords; ++i)
	{
	  rtx targ_piece = operand_subword (target, i, 1, mode);
	  rtx op0_piece = operand_subword_force (op0, i, mode);

	  if (i == word)
	    {
	      temp = expand_binop (imode, code == ABS ? and_optab : xor_optab,
				   op0_piece,
				   immed_wide_int_const (mask, imode),
				   targ_piece, 1, OPTAB_LIB_WIDEN);
	      if (temp != targ_piece)
		emit_move_insn (targ_piece, temp);
	    }
	  else
	    emit_move_insn (targ_piece, op0_piece);
	}

      insns = get_insns ();
      end_sequence ();

      emit_insn (insns);
    }
  else
    {
      temp = expand_binop (imode, code == ABS ? and_optab : xor_optab,
			   gen_lowpart (imode, op0),
			   immed_wide_int_const (mask, imode),
		           gen_lowpart (imode, target), 1, OPTAB_LIB_WIDEN);
      target = lowpart_subreg_maybe_copy (mode, temp, imode);

      set_dst_reg_note (get_last_insn (), REG_EQUAL,
			gen_rtx_fmt_e (code, mode, copy_rtx (op0)),
			target);
    }

  return target;
}

/* As expand_unop, but will fail rather than attempt the operation in a
   different mode or with a libcall.  */
static rtx
expand_unop_direct (machine_mode mode, optab unoptab, rtx op0, rtx target,
		    int unsignedp)
{
  if (optab_handler (unoptab, mode) != CODE_FOR_nothing)
    {
      struct expand_operand ops[2];
      enum insn_code icode = optab_handler (unoptab, mode);
      rtx_insn *last = get_last_insn ();
      rtx_insn *pat;

      create_output_operand (&ops[0], target, mode);
      create_convert_operand_from (&ops[1], op0, mode, unsignedp);
      pat = maybe_gen_insn (icode, 2, ops);
      if (pat)
	{
	  if (INSN_P (pat) && NEXT_INSN (pat) != NULL_RTX
	      && ! add_equal_note (pat, ops[0].value,
				   optab_to_code (unoptab),
				   ops[1].value, NULL_RTX, mode))
	    {
	      delete_insns_since (last);
	      return expand_unop (mode, unoptab, op0, NULL_RTX, unsignedp);
	    }

	  emit_insn (pat);

	  return ops[0].value;
	}
    }
  return 0;
}

/* Generate code to perform an operation specified by UNOPTAB
   on operand OP0, with result having machine-mode MODE.

   UNSIGNEDP is for the case where we have to widen the operands
   to perform the operation.  It says to use zero-extension.

   If TARGET is nonzero, the value
   is generated there, if it is convenient to do so.
   In all cases an rtx is returned for the locus of the value;
   this may or may not be TARGET.  */

rtx
expand_unop (machine_mode mode, optab unoptab, rtx op0, rtx target,
	     int unsignedp)
{
  enum mode_class mclass = GET_MODE_CLASS (mode);
  machine_mode wider_mode;
  scalar_int_mode int_mode;
  scalar_float_mode float_mode;
  rtx temp;
  rtx libfunc;

  temp = expand_unop_direct (mode, unoptab, op0, target, unsignedp);
  if (temp)
    return temp;

  /* It can't be done in this mode.  Can we open-code it in a wider mode?  */

  /* Widening (or narrowing) clz needs special treatment.  */
  if (unoptab == clz_optab)
    {
      if (is_a <scalar_int_mode> (mode, &int_mode))
	{
	  temp = widen_leading (int_mode, op0, target, unoptab);
	  if (temp)
	    return temp;

	  if (GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD
	      && optab_handler (unoptab, word_mode) != CODE_FOR_nothing)
	    {
	      temp = expand_doubleword_clz (int_mode, op0, target);
	      if (temp)
		return temp;
	    }
	}

      goto try_libcall;
    }

  if (unoptab == clrsb_optab)
    {
      if (is_a <scalar_int_mode> (mode, &int_mode))
	{
	  temp = widen_leading (int_mode, op0, target, unoptab);
	  if (temp)
	    return temp;
	}
      goto try_libcall;
    }

  if (unoptab == popcount_optab
      && is_a <scalar_int_mode> (mode, &int_mode)
      && GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD
      && optab_handler (unoptab, word_mode) != CODE_FOR_nothing
      && optimize_insn_for_speed_p ())
    {
      temp = expand_doubleword_popcount (int_mode, op0, target);
      if (temp)
	return temp;
    }

  if (unoptab == parity_optab
      && is_a <scalar_int_mode> (mode, &int_mode)
      && GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD
      && (optab_handler (unoptab, word_mode) != CODE_FOR_nothing
	  || optab_handler (popcount_optab, word_mode) != CODE_FOR_nothing)
      && optimize_insn_for_speed_p ())
    {
      temp = expand_doubleword_parity (int_mode, op0, target);
      if (temp)
	return temp;
    }

  /* Widening (or narrowing) bswap needs special treatment.  */
  if (unoptab == bswap_optab)
    {
      /* HImode is special because in this mode BSWAP is equivalent to ROTATE
	 or ROTATERT.  First try these directly; if this fails, then try the
	 obvious pair of shifts with allowed widening, as this will probably
	 be always more efficient than the other fallback methods.  */
      if (mode == HImode)
	{
	  rtx_insn *last;
	  rtx temp1, temp2;

	  if (optab_handler (rotl_optab, mode) != CODE_FOR_nothing)
	    {
	      temp = expand_binop (mode, rotl_optab, op0,
				   gen_int_shift_amount (mode, 8),
				   target, unsignedp, OPTAB_DIRECT);
	      if (temp)
		return temp;
	     }

	  if (optab_handler (rotr_optab, mode) != CODE_FOR_nothing)
	    {
	      temp = expand_binop (mode, rotr_optab, op0,
				   gen_int_shift_amount (mode, 8),
				   target, unsignedp, OPTAB_DIRECT);
	      if (temp)
		return temp;
	    }

	  last = get_last_insn ();

	  temp1 = expand_binop (mode, ashl_optab, op0,
				gen_int_shift_amount (mode, 8), NULL_RTX,
			        unsignedp, OPTAB_WIDEN);
	  temp2 = expand_binop (mode, lshr_optab, op0,
				gen_int_shift_amount (mode, 8), NULL_RTX,
			        unsignedp, OPTAB_WIDEN);
	  if (temp1 && temp2)
	    {
	      temp = expand_binop (mode, ior_optab, temp1, temp2, target,
				   unsignedp, OPTAB_WIDEN);
	      if (temp)
		return temp;
	    }

	  delete_insns_since (last);
	}

      if (is_a <scalar_int_mode> (mode, &int_mode))
	{
	  temp = widen_bswap (int_mode, op0, target);
	  if (temp)
	    return temp;

	  if (GET_MODE_SIZE (int_mode) == 2 * UNITS_PER_WORD
	      && optab_handler (unoptab, word_mode) != CODE_FOR_nothing)
	    {
	      temp = expand_doubleword_bswap (mode, op0, target);
	      if (temp)
		return temp;
	    }
	}

      goto try_libcall;
    }

  if (CLASS_HAS_WIDER_MODES_P (mclass))
    FOR_EACH_WIDER_MODE (wider_mode, mode)
      {
	if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing)
	  {
	    rtx xop0 = op0;
	    rtx_insn *last = get_last_insn ();

	    /* For certain operations, we need not actually extend
	       the narrow operand, as long as we will truncate the
	       results to the same narrowness.  */

	    xop0 = widen_operand (xop0, wider_mode, mode, unsignedp,
				  (unoptab == neg_optab
				   || unoptab == one_cmpl_optab)
				  && mclass == MODE_INT);

	    temp = expand_unop (wider_mode, unoptab, xop0, NULL_RTX,
				unsignedp);

	    if (temp)
	      {
		if (mclass != MODE_INT
		    || !TRULY_NOOP_TRUNCATION_MODES_P (mode, wider_mode))
		  {
		    if (target == 0)
		      target = gen_reg_rtx (mode);
		    convert_move (target, temp, 0);
		    return target;
		  }
		else
		  return gen_lowpart (mode, temp);
	      }
	    else
	      delete_insns_since (last);
	  }
      }

  /* These can be done a word at a time.  */
  if (unoptab == one_cmpl_optab
      && is_int_mode (mode, &int_mode)
      && GET_MODE_SIZE (int_mode) > UNITS_PER_WORD
      && optab_handler (unoptab, word_mode) != CODE_FOR_nothing)
    {
      int i;
      rtx_insn *insns;

      if (target == 0
	  || target == op0
	  || reg_overlap_mentioned_p (target, op0)
	  || !valid_multiword_target_p (target))
	target = gen_reg_rtx (int_mode);

      start_sequence ();

      /* Do the actual arithmetic.  */
      for (i = 0; i < GET_MODE_BITSIZE (int_mode) / BITS_PER_WORD; i++)
	{
	  rtx target_piece = operand_subword (target, i, 1, int_mode);
	  rtx x = expand_unop (word_mode, unoptab,
			       operand_subword_force (op0, i, int_mode),
			       target_piece, unsignedp);

	  if (target_piece != x)
	    emit_move_insn (target_piece, x);
	}

      insns = get_insns ();
      end_sequence ();

      emit_insn (insns);
      return target;
    }

  if (optab_to_code (unoptab) == NEG)
    {
      /* Try negating floating point values by flipping the sign bit.  */
      if (is_a <scalar_float_mode> (mode, &float_mode))
	{
	  temp = expand_absneg_bit (NEG, float_mode, op0, target);
	  if (temp)
	    return temp;
	}

      /* If there is no negation pattern, and we have no negative zero,
	 try subtracting from zero.  */
      if (!HONOR_SIGNED_ZEROS (mode))
	{
	  temp = expand_binop (mode, (unoptab == negv_optab
				      ? subv_optab : sub_optab),
			       CONST0_RTX (mode), op0, target,
			       unsignedp, OPTAB_DIRECT);
	  if (temp)
	    return temp;
	}
    }

  /* Try calculating parity (x) as popcount (x) % 2.  */
  if (unoptab == parity_optab && is_a <scalar_int_mode> (mode, &int_mode))
    {
      temp = expand_parity (int_mode, op0, target);
      if (temp)
	return temp;
    }

  /* Try implementing ffs (x) in terms of clz (x).  */
  if (unoptab == ffs_optab && is_a <scalar_int_mode> (mode, &int_mode))
    {
      temp = expand_ffs (int_mode, op0, target);
      if (temp)
	return temp;
    }

  /* Try implementing ctz (x) in terms of clz (x).  */
  if (unoptab == ctz_optab && is_a <scalar_int_mode> (mode, &int_mode))
    {
      temp = expand_ctz (int_mode, op0, target);
      if (temp)
	return temp;
    }

 try_libcall:
  /* Now try a library call in this mode.  */
  libfunc = optab_libfunc (unoptab, mode);
  if (libfunc)
    {
      rtx_insn *insns;
      rtx value;
      rtx eq_value;
      machine_mode outmode = mode;

      /* All of these functions return small values.  Thus we choose to
	 have them return something that isn't a double-word.  */
      if (unoptab == ffs_optab || unoptab == clz_optab || unoptab == ctz_optab
	  || unoptab == clrsb_optab || unoptab == popcount_optab
	  || unoptab == parity_optab)
	outmode
	  = GET_MODE (hard_libcall_value (TYPE_MODE (integer_type_node),
					  optab_libfunc (unoptab, mode)));

      start_sequence ();

      /* Pass 1 for NO_QUEUE so we don't lose any increments
	 if the libcall is cse'd or moved.  */
      value = emit_library_call_value (libfunc, NULL_RTX, LCT_CONST, outmode,
				       op0, mode);
      insns = get_insns ();
      end_sequence ();

      target = gen_reg_rtx (outmode);
      bool trapv = trapv_unoptab_p (unoptab);
      if (trapv)
	eq_value = NULL_RTX;
      else
	{
	  eq_value = gen_rtx_fmt_e (optab_to_code (unoptab), mode, op0);
	  if (GET_MODE_UNIT_SIZE (outmode) < GET_MODE_UNIT_SIZE (mode))
	    eq_value = simplify_gen_unary (TRUNCATE, outmode, eq_value, mode);
	  else if (GET_MODE_UNIT_SIZE (outmode) > GET_MODE_UNIT_SIZE (mode))
	    eq_value = simplify_gen_unary (ZERO_EXTEND,
					   outmode, eq_value, mode);
	}
      emit_libcall_block_1 (insns, target, value, eq_value, trapv);

      return target;
    }

  /* It can't be done in this mode.  Can we do it in a wider mode?  */

  if (CLASS_HAS_WIDER_MODES_P (mclass))
    {
      FOR_EACH_WIDER_MODE (wider_mode, mode)
	{
	  if (optab_handler (unoptab, wider_mode) != CODE_FOR_nothing
	      || optab_libfunc (unoptab, wider_mode))
	    {
	      rtx xop0 = op0;
	      rtx_insn *last = get_last_insn ();

	      /* For certain operations, we need not actually extend
		 the narrow operand, as long as we will truncate the
		 results to the same narrowness.  */
	      xop0 = widen_operand (xop0, wider_mode, mode, unsignedp,
				    (unoptab == neg_optab
				     || unoptab == one_cmpl_optab
				     || unoptab == bswap_optab)
				    && mclass == MODE_INT);

	      temp = expand_unop (wider_mode, unoptab, xop0, NULL_RTX,
				  unsignedp);

	      /* If we are generating clz using wider mode, adjust the
		 result.  Similarly for clrsb.  */
	      if ((unoptab == clz_optab || unoptab == clrsb_optab)
		  && temp != 0)
		{
		  scalar_int_mode wider_int_mode
		    = as_a <scalar_int_mode> (wider_mode);
		  int_mode = as_a <scalar_int_mode> (mode);
		  temp = expand_binop
		    (wider_mode, sub_optab, temp,
		     gen_int_mode (GET_MODE_PRECISION (wider_int_mode)
				   - GET_MODE_PRECISION (int_mode),
				   wider_int_mode),
		     target, true, OPTAB_DIRECT);
		}

	      /* Likewise for bswap.  */
	      if (unoptab == bswap_optab && temp != 0)
		{
		  scalar_int_mode wider_int_mode
		    = as_a <scalar_int_mode> (wider_mode);
		  int_mode = as_a <scalar_int_mode> (mode);
		  gcc_assert (GET_MODE_PRECISION (wider_int_mode)
			      == GET_MODE_BITSIZE (wider_int_mode)
			      && GET_MODE_PRECISION (int_mode)
				 == GET_MODE_BITSIZE (int_mode));

		  temp = expand_shift (RSHIFT_EXPR, wider_int_mode, temp,
				       GET_MODE_BITSIZE (wider_int_mode)
				       - GET_MODE_BITSIZE (int_mode),
				       NULL_RTX, true);
		}

	      if (temp)
		{