public inbox for gcc-patches@gcc.gnu.org
 help / color / mirror / Atom feed
* [ping][patch] Fixed-point patch 2/10
@ 2007-08-01 23:24 Fu, Chao-Ying
  2007-08-06  3:50 ` Mark Mitchell
  0 siblings, 1 reply; 8+ messages in thread
From: Fu, Chao-Ying @ 2007-08-01 23:24 UTC (permalink / raw)
  To: gcc-patches; +Cc: Thekkath, Radhika, Stephens, Nigel, Mark Mitchell

[-- Attachment #1: Type: text/plain, Size: 1570 bytes --]

Hi,

  Since there is no response, I re-sent the fixed-point 
patch #2.  The patch #3 to #10 will be sent out soon.
(Note that these patches are managed by Quilt.)
Can maintainers review these patches?  Thanks a lot!

> 1. Merge in machine modes to support signed and unsigned
> fract and accum modes.  Handle scalar and vector modes.
DONE!
> 2. Merge in fixed-value.h and fixed-value.c to handle fixed-point values.
(fixed-value.diff)
> 3. Merge in TREE structures for fixed-point types and constants.
> 4. Merge in C front-end changes to parse _Sat, _Fract and _Accum.
> 5. Merge in RTL structures for fixed-point constants and operators.
> 6. Merge in libcpp to parse fixed-point constants.
> 7. Merge in changes to support "case" of FIXED_POINT_TYPE, FIXED_CST,
> and CONST_FIXED in .c and .h files.
> 8. Merge in the MIPS backend that supports fixed-point instructions.
> 9. Merge in configure/build system changes for the runtime library.
> 10. Merge in testsuite (from gcc.dg/fixed-point) that only run
> when the compiler is configured to enable fixed-point.

Regards,
Chao-ying

gcc/ChangeLog
2007-08-01  Chao-ying Fu  <fu@mips.com>

        * fixed-value.h: New file.
        * fixed-value.c: New file.
        * Makefile.in (OBJS-common): Add fixed-value.o.
        (fixed-value.o): New rule.
        (GTFILES): Add fixed-value.h.
        * double-int.c (double_int_scmp): Use casts of unsigned HOST_WIDE_INT
        to compare a.low and b.low.
        * gengtype.c (main): Handle FIXED_VALUE_TYPE type as scalar typedef.

[-- Attachment #2: fixed-value.diff --]
[-- Type: application/octet-stream, Size: 39904 bytes --]

Index: gcc4x/gcc/gcc/fixed-value.h
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ gcc4x/gcc/gcc/fixed-value.h	2007-08-01 15:55:26.000000000 -0700
@@ -0,0 +1,100 @@
+/* Fixed-point arithmetic support.
+   Copyright (C) 2006, 2007 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/>.  */
+
+#ifndef GCC_FIXED_VALUE_H
+#define GCC_FIXED_VALUE_H
+
+#include "machmode.h"
+#include "real.h"
+#include "double-int.h"
+
+struct fixed_value GTY(())
+{
+  double_int data;	/* Store data up to 2 wide integers.  */
+  unsigned int mode;	/* Use machine mode to know IBIT and FBIT.  */
+};
+
+#define FIXED_VALUE_TYPE struct fixed_value
+
+#define MAX_FCONST0	18	/* For storing 18 fixed-point zeros per
+				   fract, ufract, accum, and uaccum modes .  */
+#define MAX_FCONST1	8	/* For storing 8 fixed-point ones per accum
+				   and uaccum modes.  */
+/* Constant fixed-point values 0 and 1.  */
+extern FIXED_VALUE_TYPE fconst0[MAX_FCONST0];
+extern FIXED_VALUE_TYPE fconst1[MAX_FCONST1];
+
+/* Macros to access fconst0 and fconst1 via machine modes.  */
+#define FCONST0(mode)	fconst0[mode - QQmode]
+#define FCONST1(mode)	fconst1[mode - HAmode]
+
+/* Return a CONST_FIXED with value R and mode M.  */
+#define CONST_FIXED_FROM_FIXED_VALUE(r, m) \
+  const_fixed_from_fixed_value (r, m)
+extern rtx const_fixed_from_fixed_value (FIXED_VALUE_TYPE, enum machine_mode);
+
+/* Initialize from a decimal or hexadecimal string.  */
+extern void fixed_from_string (FIXED_VALUE_TYPE *, const char *,
+			       enum machine_mode);
+
+/* In tree.c: wrap up a FIXED_VALUE_TYPE in a tree node.  */
+extern tree build_fixed (tree, FIXED_VALUE_TYPE);
+
+/* Extend or truncate to a new mode.  */
+extern bool fixed_convert (FIXED_VALUE_TYPE *, enum machine_mode,
+			   const FIXED_VALUE_TYPE *, bool);
+
+/* Convert to a fixed-point mode from an integer.  */
+extern bool fixed_convert_from_int (FIXED_VALUE_TYPE *, enum machine_mode,
+				    double_int, bool, bool);
+
+/* Convert to a fixed-point mode from a real.  */
+extern bool fixed_convert_from_real (FIXED_VALUE_TYPE *, enum machine_mode,
+				     const REAL_VALUE_TYPE *, bool);
+
+/* Convert to a real mode from a fixed-point.  */
+extern void real_convert_from_fixed (REAL_VALUE_TYPE *, enum machine_mode,
+				     const FIXED_VALUE_TYPE *);
+
+/* Compare two fixed-point objects for bitwise identity.  */
+extern bool fixed_identical (const FIXED_VALUE_TYPE *, const FIXED_VALUE_TYPE *);
+
+/* Calculate a hash value.  */
+extern unsigned int fixed_hash (const FIXED_VALUE_TYPE *);
+
+#define FIXED_VALUES_IDENTICAL(x, y)	fixed_identical (&(x), &(y))
+
+/* Determine whether a fixed-point value X is negative.  */
+#define FIXED_VALUE_NEGATIVE(x)		fixed_isneg (&(x))
+
+/* Render F as a decimal floating point constant.  */
+extern void fixed_to_decimal (char *str, const FIXED_VALUE_TYPE *, size_t);
+
+/* Binary or unary arithmetic on tree_code.  */
+extern bool fixed_arithmetic (FIXED_VALUE_TYPE *, int, const FIXED_VALUE_TYPE *,
+			      const FIXED_VALUE_TYPE *, bool);
+
+/* Compare fixed-point values by tree_code.  */
+extern bool fixed_compare (int, const FIXED_VALUE_TYPE *,
+			   const FIXED_VALUE_TYPE *);
+
+/* Determine whether a fixed-point value X is negative.  */
+extern bool fixed_isneg (const FIXED_VALUE_TYPE *);
+
+#endif /* GCC_FIXED_VALUE_H */
Index: gcc4x/gcc/gcc/fixed-value.c
===================================================================
--- /dev/null	1970-01-01 00:00:00.000000000 +0000
+++ gcc4x/gcc/gcc/fixed-value.c	2007-08-01 15:55:35.000000000 -0700
@@ -0,0 +1,1140 @@
+/* Fixed-point arithmetic support.
+   Copyright (C) 2006, 2007 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 "tm.h"
+#include "tree.h"
+#include "toplev.h"
+#include "fixed-value.h"
+
+/* Compare two fixed objects for bitwise identity.  */
+
+bool
+fixed_identical (const FIXED_VALUE_TYPE *a, const FIXED_VALUE_TYPE *b)
+{
+  return (a->mode == b->mode
+	  && a->data.high == b->data.high
+	  && a->data.low == b->data.low);
+}
+
+/* Calculate a hash value.  */
+
+unsigned int
+fixed_hash (const FIXED_VALUE_TYPE *f)
+{
+  return (unsigned int) (f->data.low ^ f->data.high);
+}
+
+/* Define the enum code for the range of the fixed-point value.  */
+enum fixed_value_range_code {
+  FIXED_OK,		/* The value is within the range.  */
+  FIXED_UNDERFLOW,	/* The value is less than the minimum.  */
+  FIXED_GT_MAX_EPS,	/* The value is greater than the maximum, but not equal
+			   to the maximum plus the epsilon.  */
+  FIXED_MAX_EPS		/* The value equals the maximum plus the epsilon.  */
+};
+
+/* Check REAL_VALUE against the range of the fixed-point mode.
+   Return FIXED_OK, if it is within the range.
+          FIXED_UNDERFLOW, if it is less than the minimum.
+          FIXED_GT_MAX_EPS, if it is greater than the maximum, but not equal to
+	    the maximum plus the epsilon.
+          FIXED_MAX_EPS, if it is equal to the maximum plus the epsilon.  */
+
+static enum fixed_value_range_code
+check_real_for_fixed_mode (REAL_VALUE_TYPE *real_value, enum machine_mode mode)
+{
+  REAL_VALUE_TYPE max_value, min_value, epsilon_value;
+
+  real_2expN (&max_value, GET_MODE_IBIT (mode));
+  real_2expN (&epsilon_value, -GET_MODE_FBIT (mode));
+
+  if (SIGNED_FIXED_POINT_MODE_P (mode))
+    min_value = REAL_VALUE_NEGATE (max_value);
+  else
+    real_from_string (&min_value, "0.0");
+
+  if (real_compare (LT_EXPR, real_value, &min_value))
+    return FIXED_UNDERFLOW;
+  if (real_compare (EQ_EXPR, real_value, &max_value))
+    return FIXED_MAX_EPS;
+  real_arithmetic (&max_value, MINUS_EXPR, &max_value, &epsilon_value);
+  if (real_compare (GT_EXPR, real_value, &max_value))
+    return FIXED_GT_MAX_EPS;
+  return FIXED_OK;
+}
+
+/* Initialize from a decimal or hexadecimal string.  */
+
+void
+fixed_from_string (FIXED_VALUE_TYPE *f, const char *str, enum machine_mode mode)
+{
+  REAL_VALUE_TYPE real_value, fixed_value, base_value;
+  unsigned int fbit;
+  enum fixed_value_range_code temp;
+
+  f->mode = mode;
+  fbit = GET_MODE_FBIT (mode);
+
+  real_from_string (&real_value, str);
+  temp = check_real_for_fixed_mode (&real_value, f->mode);
+  /* We don't want to warn the case when the _Fract value is 1.0.  */
+  if (temp == FIXED_UNDERFLOW
+      || temp == FIXED_GT_MAX_EPS
+      || (temp == FIXED_MAX_EPS && ALL_ACCUM_MODE_P (f->mode)))
+    warning (OPT_Woverflow,
+	     "large fixed-point constant implicitly truncated to fixed-point type");
+  real_2expN (&base_value, fbit);
+  real_arithmetic (&fixed_value, MULT_EXPR, &real_value, &base_value);
+  real_to_integer2 ((HOST_WIDE_INT *)&f->data.low, &f->data.high,
+		    &fixed_value);
+
+  if (temp == FIXED_MAX_EPS && ALL_FRACT_MODE_P (f->mode))
+    {
+      /* From the spec, we need to evaluate 1 to the maximal value.  */
+      f->data.low = -1;
+      f->data.high = -1;
+      f->data = double_int_ext (f->data,
+				GET_MODE_FBIT (f->mode)
+				+ GET_MODE_IBIT (f->mode), 1);
+    }
+  else
+    f->data = double_int_ext (f->data,
+			      SIGNED_FIXED_POINT_MODE_P (f->mode)
+			      + GET_MODE_FBIT (f->mode)
+			      + GET_MODE_IBIT (f->mode),
+			      UNSIGNED_FIXED_POINT_MODE_P (f->mode));
+}
+
+/* Render F as a decimal floating point constant.  */
+
+void
+fixed_to_decimal (char *str, const FIXED_VALUE_TYPE *f_orig,
+		  size_t buf_size)
+{
+  REAL_VALUE_TYPE real_value, base_value, fixed_value;
+
+  real_2expN (&base_value, GET_MODE_FBIT (f_orig->mode));
+  real_from_integer (&real_value, VOIDmode, f_orig->data.low, f_orig->data.high,
+		     UNSIGNED_FIXED_POINT_MODE_P (f_orig->mode));
+  real_arithmetic (&fixed_value, RDIV_EXPR, &real_value, &base_value);
+  real_to_decimal (str, &fixed_value, buf_size, 0, 1);
+}
+
+/* If SAT_P, saturate A to the maximum or the minimum, and save to *F based on
+   the machine mode MODE.
+   Do not modify *F otherwise.
+   This function assumes the width of double_int is greater than the width
+   of the fixed-point value (the sum of a possible sign bit, possible ibits,
+   and fbits).
+   Return true, if !SAT_P and overflow.  */
+
+static bool
+fixed_saturate1 (enum machine_mode mode, double_int a, double_int *f,
+		 bool sat_p)
+{
+  bool overflow_p = false;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (mode);
+  int i_f_bits = GET_MODE_IBIT (mode) + GET_MODE_FBIT (mode);
+
+  if (unsigned_p) /* Unsigned type.  */
+    {
+      double_int max;
+      max.low = -1;
+      max.high = -1;
+      max = double_int_ext (max, i_f_bits, 1);
+      if (double_int_cmp (a, max, 1) == 1)
+	{
+	  if (sat_p)
+	    *f = max;
+	  else
+	    overflow_p = true;
+	}
+    }
+  else /* Signed type.  */
+    {
+      double_int max, min;
+      max.high = -1;
+      max.low = -1;
+      max = double_int_ext (max, i_f_bits, 1);
+      min.high = 0;
+      min.low = 1;
+      lshift_double (min.low, min.high, i_f_bits,
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &min.low, &min.high, 1);
+      min = double_int_ext (min, 1 + i_f_bits, 0);
+      if (double_int_cmp (a, max, 0) == 1)
+	{
+	  if (sat_p)
+	    *f = max;
+	  else
+	    overflow_p = true;
+	}
+      else if (double_int_cmp (a, min, 0) == -1)
+	{
+	  if (sat_p)
+	    *f = min;
+	  else
+	    overflow_p = true;
+	}
+    }
+  return overflow_p;
+}
+
+/* If SAT_P, saturate {A_HIGH, A_LOW} to the maximum or the minimum, and
+   save to *F based on the machine mode MODE.
+   Do not modify *F otherwise.
+   This function assumes the width of two double_int is greater than the width
+   of the fixed-point value (the sum of a possible sign bit, possible ibits,
+   and fbits).
+   Return true, if !SAT_P and overflow.  */
+
+static bool
+fixed_saturate2 (enum machine_mode mode, double_int a_high, double_int a_low,
+		 double_int *f, bool sat_p)
+{
+  bool overflow_p = false;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (mode);
+  int i_f_bits = GET_MODE_IBIT (mode) + GET_MODE_FBIT (mode);
+
+  if (unsigned_p) /* Unsigned type.  */
+    {
+      double_int max_r, max_s;
+      max_r.high = 0;
+      max_r.low = 0;
+      max_s.high = -1;
+      max_s.low = -1;
+      max_s = double_int_ext (max_s, i_f_bits, 1);
+      if (double_int_cmp (a_high, max_r, 1) == 1
+	  || (double_int_equal_p (a_high, max_r) &&
+	      double_int_cmp (a_low, max_s, 1) == 1))
+	{
+	  if (sat_p)
+	    *f = max_s;
+	  else
+	    overflow_p = true;
+	}
+    }
+  else /* Signed type.  */
+    {
+      double_int max_r, max_s, min_r, min_s;
+      max_r.high = 0;
+      max_r.low = 0;
+      max_s.high = -1;
+      max_s.low = -1;
+      max_s = double_int_ext (max_s, i_f_bits, 1);
+      min_r.high = -1;
+      min_r.low = -1;
+      min_s.high = 0;
+      min_s.low = 1;
+      lshift_double (min_s.low, min_s.high, i_f_bits,
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &min_s.low, &min_s.high, 1);
+      min_s = double_int_ext (min_s, 1 + i_f_bits, 0);
+      if (double_int_cmp (a_high, max_r, 0) == 1
+	  || (double_int_equal_p (a_high, max_r) &&
+	      double_int_cmp (a_low, max_s, 1) == 1))
+	{
+	  if (sat_p)
+	    *f = max_s;
+	  else
+	    overflow_p = true;
+	}
+      else if (double_int_cmp (a_high, min_r, 0) == -1
+	       || (double_int_equal_p (a_high, min_r) &&
+		   double_int_cmp (a_low, min_s, 1) == -1))
+	{
+	  if (sat_p)
+	    *f = min_s;
+	  else
+	    overflow_p = true;
+	}
+    }
+  return overflow_p;
+}
+
+/* Return the sign bit based on I_F_BITS.  */
+
+static inline int
+get_fixed_sign_bit (double_int a, int i_f_bits)
+{
+  if (i_f_bits < HOST_BITS_PER_WIDE_INT)
+    return (a.low >> i_f_bits) & 1;
+  else
+    return (a.high >> (i_f_bits - HOST_BITS_PER_WIDE_INT)) & 1;
+}
+
+/* Calculate F = A + (SUBTRACT_P ? -B : B).
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+static bool
+do_fixed_add (FIXED_VALUE_TYPE *f, const FIXED_VALUE_TYPE *a,
+	      const FIXED_VALUE_TYPE *b, bool subtract_p, bool sat_p)
+{
+  bool overflow_p = false;
+  double_int temp = subtract_p ? double_int_neg (b->data) : b->data;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (a->mode);
+  int i_f_bits = GET_MODE_IBIT (a->mode) + GET_MODE_FBIT (a->mode);
+  f->mode = a->mode;
+  f->data = double_int_add (a->data, temp);
+  if (unsigned_p) /* Unsigned type.  */
+    {
+      if (subtract_p) /* Unsigned subtraction.  */
+	{
+	  if (double_int_cmp (a->data, b->data, 1) == -1)
+	    {
+	      if (sat_p)
+		{
+		  f->data.high = 0;
+		  f->data.low = 0;
+		 }
+	      else
+		overflow_p = true;
+	    }
+	}
+      else /* Unsigned addition.  */
+	{
+	  f->data = double_int_ext (f->data, i_f_bits, 1);
+	  if (double_int_cmp (f->data, a->data, 1) == -1
+	      || double_int_cmp (f->data, b->data, 1) == -1)
+	    {
+	      if (sat_p)
+		{
+		  f->data.high = -1;
+		  f->data.low = -1;
+		}
+	      else
+		overflow_p = true;
+	    }
+	}
+    }
+  else /* Signed type.  */
+    {
+      if ((!subtract_p
+	   && (get_fixed_sign_bit (a->data, i_f_bits)
+	       == get_fixed_sign_bit (b->data, i_f_bits))
+	   && (get_fixed_sign_bit (a->data, i_f_bits)
+	       != get_fixed_sign_bit (f->data, i_f_bits)))
+	  || (subtract_p
+	      && (get_fixed_sign_bit (a->data, i_f_bits)
+		  != get_fixed_sign_bit (b->data, i_f_bits))
+	      && (get_fixed_sign_bit (a->data, i_f_bits)
+		  != get_fixed_sign_bit (f->data, i_f_bits))))
+	{
+	  if (sat_p)
+	    {
+	      f->data.low = 1;
+	      f->data.high = 0;
+	      lshift_double (f->data.low, f->data.high, i_f_bits,
+			     2 * HOST_BITS_PER_WIDE_INT,
+			     &f->data.low, &f->data.high, 1);
+	      if (get_fixed_sign_bit (a->data, i_f_bits) == 0)
+		{
+		  double_int one;
+		  one.low = 1;
+		  one.high = 0;
+		  f->data = double_int_add (f->data, double_int_neg (one));
+		}
+	    }
+	  else
+	    overflow_p = true;
+	}
+    }
+  f->data = double_int_ext (f->data, (!unsigned_p) + i_f_bits, unsigned_p);
+  return overflow_p;
+}
+
+/* Calculate F = A * B.
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+static bool
+do_fixed_multiply (FIXED_VALUE_TYPE *f, const FIXED_VALUE_TYPE *a,
+		   const FIXED_VALUE_TYPE *b, bool sat_p)
+{
+  bool overflow_p = false;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (a->mode);
+  int i_f_bits = GET_MODE_IBIT (a->mode) + GET_MODE_FBIT (a->mode);
+  f->mode = a->mode;
+  if (GET_MODE_PRECISION (f->mode) <= HOST_BITS_PER_WIDE_INT)
+    {
+      f->data = double_int_mul (a->data, b->data);
+      lshift_double (f->data.low, f->data.high,
+		     (-GET_MODE_FBIT (f->mode)),
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &f->data.low, &f->data.high, !unsigned_p);
+      overflow_p = fixed_saturate1 (f->mode, f->data, &f->data, sat_p);
+    }
+  else
+    {
+      /* The result of multiplication expands to two double_int.  */
+      double_int a_high, a_low, b_high, b_low;
+      double_int high_high, high_low, low_high, low_low;
+      double_int r, s, temp1, temp2;
+      int carry = 0;
+
+      /* Decompose a and b to four double_int.  */
+      a_high.low = a->data.high;
+      a_high.high = 0;
+      a_low.low = a->data.low;
+      a_low.high = 0;
+      b_high.low = b->data.high;
+      b_high.high = 0;
+      b_low.low = b->data.low;
+      b_low.high = 0;
+
+      /* Perform four multiplications.  */
+      low_low = double_int_mul (a_low, b_low);
+      low_high = double_int_mul (a_low, b_high);
+      high_low = double_int_mul (a_high, b_low);
+      high_high = double_int_mul (a_high, b_high);
+
+      /* Accumulate four results to {r, s}.  */
+      temp1.high = high_low.low;
+      temp1.low = 0;
+      s = double_int_add (low_low, temp1);
+      if (double_int_cmp (s, low_low, 1) == -1
+	  || double_int_cmp (s, temp1, 1) == -1)
+	carry ++; /* Carry */
+      temp1.high = s.high;
+      temp1.low = s.low;
+      temp2.high = low_high.low;
+      temp2.low = 0;
+      s = double_int_add (temp1, temp2);
+      if (double_int_cmp (s, temp1, 1) == -1
+	  || double_int_cmp (s, temp2, 1) == -1)
+	carry ++; /* Carry */
+
+      temp1.low = high_low.high;
+      temp1.high = 0;
+      r = double_int_add (high_high, temp1);
+      temp1.low = low_high.high;
+      temp1.high = 0;
+      r = double_int_add (r, temp1);
+      temp1.low = carry;
+      temp1.high = 0;
+      r = double_int_add (r, temp1);
+
+      /* We need to add neg(b) to r, if a < 0.  */
+      if (!unsigned_p && a->data.high < 0)
+	r = double_int_add (r, double_int_neg (b->data));
+      /* We need to add neg(a) to r, if b < 0.  */
+      if (!unsigned_p && b->data.high < 0)
+	r = double_int_add (r, double_int_neg (a->data));
+
+      /* Shift right the result by FBIT.  */
+      if (GET_MODE_FBIT (f->mode) == 2 * HOST_BITS_PER_WIDE_INT)
+	{
+	  s.low = r.low;
+	  s.high = r.high;
+	  if (unsigned_p)
+	    {
+	      r.low = 0;
+	      r.high = 0;
+	    }
+	  else
+	    {
+	      r.low = -1;
+	      r.high = -1;
+	    }
+	  f->data.low = s.low;
+	  f->data.high = s.high;
+	}
+      else
+	{
+	  lshift_double (s.low, s.high,
+			 (-GET_MODE_FBIT (f->mode)),
+			 2 * HOST_BITS_PER_WIDE_INT,
+			 &s.low, &s.high, 0);
+	  lshift_double (r.low, r.high,
+			 (2 * HOST_BITS_PER_WIDE_INT
+			  - GET_MODE_FBIT (f->mode)),
+			 2 * HOST_BITS_PER_WIDE_INT,
+			 &f->data.low, &f->data.high, 0);
+	  f->data.low = f->data.low | s.low;
+	  f->data.high = f->data.high | s.high;
+	  s.low = f->data.low;
+	  s.high = f->data.high;
+	  lshift_double (r.low, r.high,
+			 (-GET_MODE_FBIT (f->mode)),
+			 2 * HOST_BITS_PER_WIDE_INT,
+			 &r.low, &r.high, !unsigned_p);
+	}
+
+      overflow_p = fixed_saturate2 (f->mode, r, s, &f->data, sat_p);
+    }
+
+  f->data = double_int_ext (f->data, (!unsigned_p) + i_f_bits, unsigned_p);
+  return overflow_p;
+}
+
+/* Calculate F = A / B.
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+static bool
+do_fixed_divide (FIXED_VALUE_TYPE *f, const FIXED_VALUE_TYPE *a,
+		 const FIXED_VALUE_TYPE *b, bool sat_p)
+{
+  bool overflow_p = false;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (a->mode);
+  int i_f_bits = GET_MODE_IBIT (a->mode) + GET_MODE_FBIT (a->mode);
+  f->mode = a->mode;
+  if (GET_MODE_PRECISION (f->mode) <= HOST_BITS_PER_WIDE_INT)
+    {
+      lshift_double (a->data.low, a->data.high,
+		     GET_MODE_FBIT (f->mode),
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &f->data.low, &f->data.high, !unsigned_p);
+      f->data = double_int_div (f->data, b->data, unsigned_p, TRUNC_DIV_EXPR);
+      overflow_p = fixed_saturate1 (f->mode, f->data, &f->data, sat_p);
+    }
+  else
+    {
+      double_int pos_a, pos_b, r, s;
+      double_int quo_r, quo_s, mod, temp;
+      int num_of_neg = 0;
+      int i;
+
+      /* If a < 0, negate a.  */
+      if (!unsigned_p && a->data.high < 0)
+	{
+	  pos_a = double_int_neg (a->data);
+	  num_of_neg ++;
+	}
+      else
+	pos_a = a->data;
+
+      /* If b < 0, negate b.  */
+      if (!unsigned_p && b->data.high < 0)
+	{
+	  pos_b = double_int_neg (b->data);
+	  num_of_neg ++;
+	}
+      else
+	pos_b = b->data;
+
+      /* Left shift pos_a to {r, s} by FBIT.  */
+      if (GET_MODE_FBIT (f->mode) == 2 * HOST_BITS_PER_WIDE_INT)
+	{
+	  r = pos_a;
+	  s.high = 0;
+	  s.low = 0;
+	}
+      else
+ 	{
+	  lshift_double (pos_a.low, pos_a.high,
+			 GET_MODE_FBIT (f->mode),
+			 2 * HOST_BITS_PER_WIDE_INT,
+			 &s.low, &s.high, 0);
+	  lshift_double (pos_a.low, pos_a.high,
+			 - (2 * HOST_BITS_PER_WIDE_INT
+			    - GET_MODE_FBIT (f->mode)),
+			 2 * HOST_BITS_PER_WIDE_INT,
+			 &r.low, &r.high, 0);
+ 	}
+
+      /* Divide r by pos_b to quo_r.  The remanider is in mod.  */
+      div_and_round_double (TRUNC_DIV_EXPR, 1, r.low, r.high, pos_b.low,
+			    pos_b.high, &quo_r.low, &quo_r.high, &mod.low,
+			    &mod.high);
+
+      quo_s.high = 0;
+      quo_s.low = 0;
+
+      for (i = 0; i < 2 * HOST_BITS_PER_WIDE_INT; i++)
+	{
+	  /* Record the leftmost bit of mod.  */
+	  int leftmost_mod = (mod.high < 0);
+
+	  /* Shift left mod by 1 bit.  */
+	  lshift_double (mod.low, mod.high, 1, 2 * HOST_BITS_PER_WIDE_INT,
+			 &mod.low, &mod.high, 0);
+
+	  /* Test the leftmost bit of s to add to mod.  */
+	  if (s.high < 0)
+	    mod.low += 1;
+
+	  /* Shift left quo_s by 1 bit.  */
+	  lshift_double (quo_s.low, quo_s.high, 1, 2 * HOST_BITS_PER_WIDE_INT,
+			 &quo_s.low, &quo_s.high, 0);
+
+	  /* Try to calculate (mod - pos_b).  */
+	  temp = double_int_add (mod, double_int_neg (pos_b));
+
+	  if (leftmost_mod == 1 || double_int_cmp (mod, pos_b, 1) != -1)
+	    {
+	      quo_s.low += 1;
+	      mod = temp;
+	    }
+
+	  /* Shift left s by 1 bit.  */
+	  lshift_double (s.low, s.high, 1, 2 * HOST_BITS_PER_WIDE_INT,
+			 &s.low, &s.high, 0);
+
+	}
+
+      if (num_of_neg == 1)
+	{
+	  quo_s = double_int_neg (quo_s);
+	  if (quo_s.high == 0 && quo_s.low == 0)
+	    quo_r = double_int_neg (quo_r);
+	  else
+	    {
+	      quo_r.low = ~quo_r.low;
+	      quo_r.high = ~quo_r.high;
+	    }
+	}
+
+      f->data = quo_s;
+      overflow_p = fixed_saturate2 (f->mode, quo_r, quo_s, &f->data, sat_p);
+    }
+
+  f->data = double_int_ext (f->data, (!unsigned_p) + i_f_bits, unsigned_p);
+  return overflow_p;
+}
+
+/* Calculate F = A << B if LEFT_P.  Otherwies, F = A >> B.
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+static bool
+do_fixed_shift (FIXED_VALUE_TYPE *f, const FIXED_VALUE_TYPE *a,
+	      const FIXED_VALUE_TYPE *b, bool left_p, bool sat_p)
+{
+  bool overflow_p = false;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (a->mode);
+  int i_f_bits = GET_MODE_IBIT (a->mode) + GET_MODE_FBIT (a->mode);
+  f->mode = a->mode;
+
+  if (b->data.low == 0)
+    {
+      f->data = a->data;
+      return overflow_p;
+    }
+
+  if (GET_MODE_PRECISION (f->mode) <= HOST_BITS_PER_WIDE_INT || (!left_p))
+    {
+      lshift_double (a->data.low, a->data.high,
+		     left_p ? b->data.low : (-b->data.low),
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &f->data.low, &f->data.high, !unsigned_p);
+      if (left_p) /* Only left shift saturates.  */
+	overflow_p = fixed_saturate1 (f->mode, f->data, &f->data, sat_p);
+    }
+  else /* We need two double_int to store the left-shift result.  */
+    {
+      double_int temp_high, temp_low;
+      if (b->data.low == 2 * HOST_BITS_PER_WIDE_INT)
+	{
+	  temp_high = a->data;
+	  temp_low.high = 0;
+	  temp_low.low = 0;
+	}
+      else
+	{
+	  lshift_double (a->data.low, a->data.high,
+			 b->data.low,
+			 2 * HOST_BITS_PER_WIDE_INT,
+			 &temp_low.low, &temp_low.high, !unsigned_p);
+	  /* Logical shift right to temp_high.  */
+	  lshift_double (a->data.low, a->data.high,
+			 b->data.low - 2 * HOST_BITS_PER_WIDE_INT,
+			 2 * HOST_BITS_PER_WIDE_INT,
+			 &temp_high.low, &temp_high.high, 0);
+	}
+      if (!unsigned_p && a->data.high < 0) /* Signed-extend temp_high.  */
+	temp_high = double_int_ext (temp_high, b->data.low, unsigned_p);
+      f->data = temp_low;
+      overflow_p = fixed_saturate2 (f->mode, temp_high, temp_low, &f->data,
+				    sat_p);
+    }
+  f->data = double_int_ext (f->data, (!unsigned_p) + i_f_bits, unsigned_p);
+  return overflow_p;
+}
+
+/* Calculate F = -A.
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+static bool
+do_fixed_neg (FIXED_VALUE_TYPE *f, const FIXED_VALUE_TYPE *a, bool sat_p)
+{
+  bool overflow_p = false;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (a->mode);
+  int i_f_bits = GET_MODE_IBIT (a->mode) + GET_MODE_FBIT (a->mode);
+  f->mode = a->mode;
+  f->data = double_int_neg (a->data);
+  f->data = double_int_ext (f->data, (!unsigned_p) + i_f_bits, unsigned_p);
+
+  if (unsigned_p) /* Unsigned type.  */
+    {
+      if (f->data.low != 0 || f->data.high != 0)
+	{
+	  if (sat_p)
+	    {
+	      f->data.low = 0;
+	      f->data.high = 0;
+	    }
+	  else
+	    overflow_p = true;
+	}
+    }
+  else /* Signed type.  */
+    {
+      if (!(f->data.high == 0 && f->data.low == 0)
+	  && f->data.high == a->data.high && f->data.low == a->data.low )
+	{
+	  if (sat_p)
+	    {
+	      /* Saturate to the maximum by subtracting f->data by one.  */
+	      f->data.low = -1;
+	      f->data.high = -1;
+	      f->data = double_int_ext (f->data, i_f_bits, 1);
+	    }
+	  else
+	    overflow_p = true;
+	}
+    }
+  return overflow_p;
+}
+
+/* Perform the binary or unary operation described by CODE.
+   Note that OP0 and OP1 must have the same mode for binary operators.
+   For a unary operation, leave OP1 NULL.
+   Return true, if !SAT_P and overflow.  */
+
+bool
+fixed_arithmetic (FIXED_VALUE_TYPE *f, int icode, const FIXED_VALUE_TYPE *op0,
+		  const FIXED_VALUE_TYPE *op1, bool sat_p)
+{
+  switch (icode)
+    {
+    case NEGATE_EXPR:
+      return do_fixed_neg (f, op0, sat_p);
+      break;
+
+    case PLUS_EXPR:
+      gcc_assert (op0->mode == op1->mode);
+      return do_fixed_add (f, op0, op1, false, sat_p);
+      break;
+
+    case MINUS_EXPR:
+      gcc_assert (op0->mode == op1->mode);
+      return do_fixed_add (f, op0, op1, true, sat_p);
+      break;
+
+    case MULT_EXPR:
+      gcc_assert (op0->mode == op1->mode);
+      return do_fixed_multiply (f, op0, op1, sat_p);
+      break;
+
+    case TRUNC_DIV_EXPR:
+      gcc_assert (op0->mode == op1->mode);
+      return do_fixed_divide (f, op0, op1, sat_p);
+      break;
+
+    case LSHIFT_EXPR:
+      return do_fixed_shift (f, op0, op1, true, sat_p);
+      break;
+
+    case RSHIFT_EXPR:
+      return do_fixed_shift (f, op0, op1, false, sat_p);
+      break;
+
+    default:
+      gcc_unreachable ();
+    }
+  return false;
+}
+
+/* Compare fixed-point values by tree_code.
+   Note that OP0 and OP1 must have the same mode.  */
+
+bool
+fixed_compare (int icode, const FIXED_VALUE_TYPE *op0,
+	       const FIXED_VALUE_TYPE *op1)
+{
+  enum tree_code code = icode;
+  gcc_assert (op0->mode == op1->mode);
+
+  switch (code)
+    {
+    case NE_EXPR:
+      return !double_int_equal_p (op0->data, op1->data);
+
+    case EQ_EXPR:
+      return double_int_equal_p (op0->data, op1->data);
+
+    case LT_EXPR:
+      return double_int_cmp (op0->data, op1->data,
+			     UNSIGNED_FIXED_POINT_MODE_P (op0->mode)) == -1;
+
+    case LE_EXPR:
+      return double_int_cmp (op0->data, op1->data,
+			     UNSIGNED_FIXED_POINT_MODE_P (op0->mode)) != 1;
+
+    case GT_EXPR:
+      return double_int_cmp (op0->data, op1->data,
+			     UNSIGNED_FIXED_POINT_MODE_P (op0->mode)) == 1;
+
+    case GE_EXPR:
+      return double_int_cmp (op0->data, op1->data,
+			     UNSIGNED_FIXED_POINT_MODE_P (op0->mode)) != -1;
+
+    default:
+      gcc_unreachable ();
+    }
+}
+
+/* Extend or truncate to a new mode.
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+bool
+fixed_convert (FIXED_VALUE_TYPE *f, enum machine_mode mode,
+               const FIXED_VALUE_TYPE *a, bool sat_p)
+{
+  bool overflow_p = false;
+  if (mode == a->mode)
+    {
+      *f = *a;
+      return overflow_p;
+    }
+
+  if (GET_MODE_FBIT (mode) > GET_MODE_FBIT (a->mode))
+    {
+      /* Left shift a to temp_high, temp_low based on a->mode.  */
+      double_int temp_high, temp_low;
+      int amount = GET_MODE_FBIT (mode) - GET_MODE_FBIT (a->mode);
+      lshift_double (a->data.low, a->data.high,
+		     amount,
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &temp_low.low, &temp_low.high,
+		     SIGNED_FIXED_POINT_MODE_P (a->mode));
+      /* Logical shift right to temp_high.  */
+      lshift_double (a->data.low, a->data.high,
+		     amount - 2 * HOST_BITS_PER_WIDE_INT,
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &temp_high.low, &temp_high.high, 0);
+      if (SIGNED_FIXED_POINT_MODE_P (a->mode)
+	  && a->data.high < 0) /* Signed-extend temp_high.  */
+	temp_high = double_int_ext (temp_high, amount, 0);
+      f->mode = mode;
+      f->data = temp_low;
+      if (SIGNED_FIXED_POINT_MODE_P (a->mode) ==
+	  SIGNED_FIXED_POINT_MODE_P (f->mode))
+	overflow_p = fixed_saturate2 (f->mode, temp_high, temp_low, &f->data,
+				      sat_p);
+      else
+	{
+	  /* Take care of the cases when converting between signed and
+	     unsigned.  */
+	  if (SIGNED_FIXED_POINT_MODE_P (a->mode))
+	    {
+	      /* Signed -> Unsigned.  */
+	      if (a->data.high < 0)
+		{
+		  if (sat_p)
+		    {
+		      f->data.low = 0;  /* Set to zero.  */
+		      f->data.high = 0;  /* Set to zero.  */
+		    }
+		  else
+		    overflow_p = true;
+		}
+	      else
+		overflow_p = fixed_saturate2 (f->mode, temp_high, temp_low,
+					      &f->data, sat_p);
+	    }
+	  else
+	    {
+	      /* Unsigned -> Signed.  */
+	      if (temp_high.high < 0)
+		{
+		  if (sat_p)
+		    {
+		      /* Set to maximum.  */
+		      f->data.low = -1;  /* Set to all ones.  */
+		      f->data.high = -1;  /* Set to all ones.  */
+		      f->data = double_int_ext (f->data,
+						GET_MODE_FBIT (f->mode)
+						+ GET_MODE_IBIT (f->mode),
+						1); /* Clear the sign.  */
+		    }
+		  else
+		    overflow_p = true;
+		}
+	      else
+		overflow_p = fixed_saturate2 (f->mode, temp_high, temp_low,
+					      &f->data, sat_p);
+	    }
+	}
+    }
+  else
+    {
+      /* Right shift a to temp based on a->mode.  */
+      double_int temp;
+      lshift_double (a->data.low, a->data.high,
+		     GET_MODE_FBIT (mode) - GET_MODE_FBIT (a->mode),
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &temp.low, &temp.high,
+		     SIGNED_FIXED_POINT_MODE_P (a->mode));
+      f->mode = mode;
+      f->data = temp;
+      if (SIGNED_FIXED_POINT_MODE_P (a->mode) ==
+	  SIGNED_FIXED_POINT_MODE_P (f->mode))
+	overflow_p = fixed_saturate1 (f->mode, f->data, &f->data, sat_p);
+      else
+	{
+	  /* Take care of the cases when converting between signed and
+	     unsigned.  */
+	  if (SIGNED_FIXED_POINT_MODE_P (a->mode))
+	    {
+	      /* Signed -> Unsigned.  */
+	      if (a->data.high < 0)
+		{
+		  if (sat_p)
+		    {
+		      f->data.low = 0;  /* Set to zero.  */
+		      f->data.high = 0;  /* Set to zero.  */
+		    }
+		  else
+		    overflow_p = true;
+		}
+	      else
+		overflow_p = fixed_saturate1 (f->mode, f->data, &f->data,
+					      sat_p);
+	    }
+	  else
+	    {
+	      /* Unsigned -> Signed.  */
+	      if (temp.high < 0)
+		{
+		  if (sat_p)
+		    {
+		      /* Set to maximum.  */
+		      f->data.low = -1;  /* Set to all ones.  */
+		      f->data.high = -1;  /* Set to all ones.  */
+		      f->data = double_int_ext (f->data,
+						GET_MODE_FBIT (f->mode)
+						+ GET_MODE_IBIT (f->mode),
+						1); /* Clear the sign.  */
+		    }
+		  else
+		    overflow_p = true;
+		}
+	      else
+		overflow_p = fixed_saturate1 (f->mode, f->data, &f->data,
+					      sat_p);
+	    }
+	}
+    }
+
+  f->data = double_int_ext (f->data,
+			    SIGNED_FIXED_POINT_MODE_P (f->mode)
+			    + GET_MODE_FBIT (f->mode)
+			    + GET_MODE_IBIT (f->mode),
+			    UNSIGNED_FIXED_POINT_MODE_P (f->mode));
+  return overflow_p;
+}
+
+/* Convert to a new fixed-point mode from an integer.
+   If UNSIGNED_P, this integer is unsigned.
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+bool
+fixed_convert_from_int (FIXED_VALUE_TYPE *f, enum machine_mode mode,
+			double_int a, bool unsigned_p, bool sat_p)
+{
+  bool overflow_p = false;
+  /* Left shift a to temp_high, temp_low.  */
+  double_int temp_high, temp_low;
+  int amount = GET_MODE_FBIT (mode);
+  if (amount == 2 * HOST_BITS_PER_WIDE_INT)
+    {
+       temp_high = a;
+       temp_low.low = 0;
+       temp_low.high = 0;
+    }
+  else
+    {
+      lshift_double (a.low, a.high,
+		     amount,
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &temp_low.low, &temp_low.high, 0);
+
+      /* Logical shift right to temp_high.  */
+      lshift_double (a.low, a.high,
+		     amount - 2 * HOST_BITS_PER_WIDE_INT,
+		     2 * HOST_BITS_PER_WIDE_INT,
+		     &temp_high.low, &temp_high.high, 0);
+    }
+  if (!unsigned_p && a.high < 0) /* Signed-extend temp_high.  */
+    temp_high = double_int_ext (temp_high, amount, 0);
+
+  f->mode = mode;
+  f->data = temp_low;
+
+  if (unsigned_p == UNSIGNED_FIXED_POINT_MODE_P (f->mode))
+    overflow_p = fixed_saturate2 (f->mode, temp_high, temp_low, &f->data,
+				  sat_p);
+  else
+    {
+      /* Take care of the cases when converting between signed and unsigned.  */
+      if (!unsigned_p)
+	{
+	  /* Signed -> Unsigned.  */
+	  if (a.high < 0)
+	    {
+	      if (sat_p)
+		{
+		  f->data.low = 0;  /* Set to zero.  */
+		  f->data.high = 0;  /* Set to zero.  */
+		}
+	      else
+		overflow_p = true;
+	    }
+	  else
+	    overflow_p = fixed_saturate2 (f->mode, temp_high, temp_low,
+					  &f->data, sat_p);
+	}
+      else
+	{
+	  /* Unsigned -> Signed.  */
+	  if (temp_high.high < 0)
+	    {
+	      if (sat_p)
+		{
+		  /* Set to maximum.  */
+		  f->data.low = -1;  /* Set to all ones.  */
+		  f->data.high = -1;  /* Set to all ones.  */
+		  f->data = double_int_ext (f->data,
+					    GET_MODE_FBIT (f->mode)
+					    + GET_MODE_IBIT (f->mode),
+					    1); /* Clear the sign.  */
+		}
+	      else
+		overflow_p = true;
+	    }
+	  else
+	    overflow_p = fixed_saturate2 (f->mode, temp_high, temp_low,
+					  &f->data, sat_p);
+	}
+    }
+  f->data = double_int_ext (f->data,
+			    SIGNED_FIXED_POINT_MODE_P (f->mode)
+			    + GET_MODE_FBIT (f->mode)
+			    + GET_MODE_IBIT (f->mode),
+			    UNSIGNED_FIXED_POINT_MODE_P (f->mode));
+  return overflow_p;
+}
+
+/* Convert to a new fixed-point mode from a real.
+   If SAT_P, saturate the result to the max or the min.
+   Return true, if !SAT_P and overflow.  */
+
+bool
+fixed_convert_from_real (FIXED_VALUE_TYPE *f, enum machine_mode mode,
+			 const REAL_VALUE_TYPE *a, bool sat_p)
+{
+  bool overflow_p = false;
+  REAL_VALUE_TYPE real_value, fixed_value, base_value;
+  bool unsigned_p = UNSIGNED_FIXED_POINT_MODE_P (mode);
+  int i_f_bits = GET_MODE_IBIT (mode) + GET_MODE_FBIT (mode);
+  unsigned int fbit = GET_MODE_FBIT (mode);
+  enum fixed_value_range_code temp;
+
+  real_value = *a;
+  f->mode = mode;
+  real_2expN (&base_value, fbit);
+  real_arithmetic (&fixed_value, MULT_EXPR, &real_value, &base_value);
+  real_to_integer2 ((HOST_WIDE_INT *)&f->data.low, &f->data.high, &fixed_value);
+  temp = check_real_for_fixed_mode (&real_value, mode);
+  if (temp == FIXED_UNDERFLOW) /* Minimum.  */
+    {
+      if (sat_p)
+	{
+	  if (unsigned_p)
+	    {
+	      f->data.low = 0;
+	      f->data.high = 0;
+	    }
+	  else
+	    {
+	      f->data.low = 1;
+	      f->data.high = 0;
+	      lshift_double (f->data.low, f->data.high, i_f_bits,
+			     2 * HOST_BITS_PER_WIDE_INT,
+			     &f->data.low, &f->data.high, 1);
+	      f->data = double_int_ext (f->data, 1 + i_f_bits, 0);
+	    }
+	}
+      else
+	overflow_p = true;
+    }
+  else if (temp == FIXED_GT_MAX_EPS || temp == FIXED_MAX_EPS) /* Maximum.  */
+    {
+      if (sat_p)
+	{
+	  f->data.low = -1;
+	  f->data.high = -1;
+	  f->data = double_int_ext (f->data, i_f_bits, 1);
+	}
+      else
+	overflow_p = true;
+    }
+  f->data = double_int_ext (f->data, (!unsigned_p) + i_f_bits, unsigned_p);
+  return overflow_p;
+}
+
+/* Convert to a new real mode from a fixed-point.  */
+
+void
+real_convert_from_fixed (REAL_VALUE_TYPE *r, enum machine_mode mode,
+			 const FIXED_VALUE_TYPE *f)
+{
+  REAL_VALUE_TYPE base_value, fixed_value, real_value;
+
+  real_2expN (&base_value, GET_MODE_FBIT (f->mode));
+  real_from_integer (&fixed_value, VOIDmode, f->data.low, f->data.high,
+		     UNSIGNED_FIXED_POINT_MODE_P (f->mode));
+  real_arithmetic (&real_value, RDIV_EXPR, &fixed_value, &base_value);
+  real_convert (r, mode, &real_value);
+}
+
+/* Determine whether a fixed-point value F is negative.  */
+
+bool
+fixed_isneg (const FIXED_VALUE_TYPE *f)
+{
+  if (SIGNED_FIXED_POINT_MODE_P (f->mode))
+    {
+      int i_f_bits = GET_MODE_IBIT (f->mode) + GET_MODE_FBIT (f->mode);
+      int sign_bit = get_fixed_sign_bit (f->data, i_f_bits);
+      if (sign_bit == 1)
+	return true;
+    }
+
+  return false;
+}
Index: gcc4x/gcc/gcc/Makefile.in
===================================================================
--- gcc4x.orig/gcc/gcc/Makefile.in	2007-08-01 15:53:53.000000000 -0700
+++ gcc4x/gcc/gcc/Makefile.in	2007-08-01 15:53:55.000000000 -0700
@@ -1028,6 +1028,7 @@
 	expmed.o \
 	expr.o \
 	final.o \
+	fixed-value.o \
 	fold-const.o \
 	function.o \
 	fwprop.o \
@@ -2413,6 +2414,8 @@
    toplev.h $(TM_P_H) $(REAL_H)
 dfp.o : dfp.c dfp.h $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H)	$(TREE_H) \
    toplev.h $(TM_P_H) $(REAL_H) $(DECNUM_H)
+fixed-value.o: fixed-value.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) \
+   $(TREE_H) fixed-value.h $(REAL_H) toplev.h
 integrate.o : integrate.c $(CONFIG_H) $(SYSTEM_H) coretypes.h $(TM_H) \
    $(RTL_H) $(TREE_H) $(FLAGS_H) debug.h $(INTEGRATE_H) insn-config.h \
    $(EXPR_H) $(REAL_H) $(REGS_H) intl.h $(FUNCTION_H) output.h $(RECOG_H) \
@@ -3038,6 +3041,7 @@
   $(srcdir)/coverage.c $(srcdir)/rtl.h \
   $(srcdir)/optabs.h $(srcdir)/tree.h $(srcdir)/function.h $(srcdir)/libfuncs.h $(SYMTAB_H) \
   $(srcdir)/real.h $(srcdir)/varray.h $(srcdir)/insn-addr.h $(srcdir)/hwint.h \
+  $(srcdir)/fixed-value.h \
   $(srcdir)/ipa-reference.h $(srcdir)/output.h $(srcdir)/cfgloop.h \
   $(srcdir)/cselib.h $(srcdir)/basic-block.h  $(srcdir)/cgraph.h \
   $(srcdir)/reload.h $(srcdir)/caller-save.c \
Index: gcc4x/gcc/gcc/double-int.c
===================================================================
--- gcc4x.orig/gcc/gcc/double-int.c	2007-08-01 15:53:53.000000000 -0700
+++ gcc4x/gcc/gcc/double-int.c	2007-08-01 15:53:55.000000000 -0700
@@ -362,9 +362,9 @@
     return -1;
   if (a.high > b.high)
     return 1;
-  if ((HOST_WIDE_INT) a.low < (HOST_WIDE_INT) b.low)
+  if ((unsigned HOST_WIDE_INT) a.low < (unsigned HOST_WIDE_INT) b.low)
     return -1;
-  if ((HOST_WIDE_INT) a.low > (HOST_WIDE_INT) b.low)
+  if ((unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low)
     return 1;
 
   return 0;
Index: gcc4x/gcc/gcc/gengtype.c
===================================================================
--- gcc4x.orig/gcc/gcc/gengtype.c	2007-08-01 15:53:53.000000000 -0700
+++ gcc4x/gcc/gcc/gengtype.c	2007-08-01 15:53:55.000000000 -0700
@@ -3533,6 +3533,7 @@
   pos.line = __LINE__ + 1;
   do_scalar_typedef ("CUMULATIVE_ARGS", &pos); pos.line++;
   do_scalar_typedef ("REAL_VALUE_TYPE", &pos); pos.line++;
+  do_scalar_typedef ("FIXED_VALUE_TYPE", &pos); pos.line++;
   do_scalar_typedef ("double_int", &pos); pos.line++;
   do_scalar_typedef ("uint8", &pos); pos.line++;
   do_scalar_typedef ("jword", &pos); pos.line++;

^ permalink raw reply	[flat|nested] 8+ messages in thread

* Re: [ping][patch] Fixed-point patch 2/10
  2007-08-01 23:24 [ping][patch] Fixed-point patch 2/10 Fu, Chao-Ying
@ 2007-08-06  3:50 ` Mark Mitchell
  2007-08-06 17:35   ` Fu, Chao-Ying
  0 siblings, 1 reply; 8+ messages in thread
From: Mark Mitchell @ 2007-08-06  3:50 UTC (permalink / raw)
  To: Fu, Chao-Ying; +Cc: gcc-patches, Thekkath, Radhika, Stephens, Nigel

Fu, Chao-Ying wrote:

> gcc/ChangeLog
> 2007-08-01  Chao-ying Fu  <fu@mips.com>
> 
>         * fixed-value.h: New file.
>         * fixed-value.c: New file.
>         * Makefile.in (OBJS-common): Add fixed-value.o.
>         (fixed-value.o): New rule.
>         (GTFILES): Add fixed-value.h.
>         * double-int.c (double_int_scmp): Use casts of unsigned HOST_WIDE_INT
>         to compare a.low and b.low.
>         * gengtype.c (main): Handle FIXED_VALUE_TYPE type as scalar typedef.

This is OK -- except that I do not understand the change to
double_int_scmp.  That function explicitly assumes that the double int
is signed, as opposed to double_int_ucmp, which assumes that it is
unsigned.  Please explain.

Thanks,

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713

^ permalink raw reply	[flat|nested] 8+ messages in thread

* RE: [ping][patch] Fixed-point patch 2/10
  2007-08-06  3:50 ` Mark Mitchell
@ 2007-08-06 17:35   ` Fu, Chao-Ying
  2007-08-06 18:08     ` Mark Mitchell
  0 siblings, 1 reply; 8+ messages in thread
From: Fu, Chao-Ying @ 2007-08-06 17:35 UTC (permalink / raw)
  To: Mark Mitchell; +Cc: gcc-patches, Thekkath, Radhika, Stephens, Nigel

Mark Mitchell wrote:
> 
> Fu, Chao-Ying wrote:
> 
> > gcc/ChangeLog
> > 2007-08-01  Chao-ying Fu  <fu@mips.com>
> > 
> >         * fixed-value.h: New file.
> >         * fixed-value.c: New file.
> >         * Makefile.in (OBJS-common): Add fixed-value.o.
> >         (fixed-value.o): New rule.
> >         (GTFILES): Add fixed-value.h.
> >         * double-int.c (double_int_scmp): Use casts of 
> unsigned HOST_WIDE_INT
> >         to compare a.low and b.low.
> >         * gengtype.c (main): Handle FIXED_VALUE_TYPE type 
> as scalar typedef.
> 
> This is OK -- except that I do not understand the change to
> double_int_scmp.  That function explicitly assumes that the double int
> is signed, as opposed to double_int_ucmp, which assumes that it is
> unsigned.  Please explain.

  I think the sign bit is in the "high" part, so we need to 
compare the "low" part as unsigned for "double_int_scmp".

Ex 1:  If A.high and B.high are 0, A and B are positive numbers.
Then, when we compare A.low and B.low, we should use unsigned compare.
A.high = 0 and A.low = 0x8000000000000000
B.high = 0 and B.low = 0x0000000000000001
=> A > B, 
because A.high == B.high and
(unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low

Ex 2: (A and B are negative. )
A.high = 0xffffffffffffffff and A.low = 0x8000000000000000
B.high = 0xffffffffffffffff and B.low = 0x0000000000000001
=> A > B, 
because A.high == B.high and
(unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low

  Thanks!

Regards,
Chao-ying

^ permalink raw reply	[flat|nested] 8+ messages in thread

* Re: [ping][patch] Fixed-point patch 2/10
  2007-08-06 17:35   ` Fu, Chao-Ying
@ 2007-08-06 18:08     ` Mark Mitchell
  2007-08-06 19:17       ` Nigel Stephens
  0 siblings, 1 reply; 8+ messages in thread
From: Mark Mitchell @ 2007-08-06 18:08 UTC (permalink / raw)
  To: Fu, Chao-Ying; +Cc: gcc-patches, Thekkath, Radhika, Stephens, Nigel

Fu, Chao-Ying wrote:

> Ex 2: (A and B are negative. )
> A.high = 0xffffffffffffffff and A.low = 0x8000000000000000
> B.high = 0xffffffffffffffff and B.low = 0x0000000000000001
> => A > B, 
> because A.high == B.high and
> (unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low

Yes, I see; two's complement is set up so that after masking the sign
bit, you can do an unsigned compare.  Thank you for explaining.

The patch is OK, then.

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713

^ permalink raw reply	[flat|nested] 8+ messages in thread

* Re: [ping][patch] Fixed-point patch 2/10
  2007-08-06 18:08     ` Mark Mitchell
@ 2007-08-06 19:17       ` Nigel Stephens
  2007-08-06 19:56         ` Mark Mitchell
  0 siblings, 1 reply; 8+ messages in thread
From: Nigel Stephens @ 2007-08-06 19:17 UTC (permalink / raw)
  To: Mark Mitchell; +Cc: Fu, Chao-Ying, gcc-patches, Thekkath, Radhika



Mark Mitchell wrote:
> Fu, Chao-Ying wrote:
>
>   
>> Ex 2: (A and B are negative. )
>> A.high = 0xffffffffffffffff and A.low = 0x8000000000000000
>> B.high = 0xffffffffffffffff and B.low = 0x0000000000000001
>> => A > B, 
>> because A.high == B.high and
>> (unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low
>>     
>
> Yes, I see; two's complement is set up so that after masking the sign
> bit, you can do an unsigned compare.  Thank you for explaining.
>   

Since double_int is already declared with low unsigned, like this:

    typedef struct
    {
      unsigned HOST_WIDE_INT low;
      HOST_WIDE_INT high;
    } double_int;


Would it be more obvious to simply remove the casts altogether, e.g.

    double_int_scmp (double_int a, double_int b)
    {
      if (a.high < b.high)
        return -1;
      if (a.high > b.high)
        return 1;
      if (a.low < b.low)
        return -1;
      if (a.low > b.low)
        return 1;

      return 0;
    }



Nigel

^ permalink raw reply	[flat|nested] 8+ messages in thread

* Re: [ping][patch] Fixed-point patch 2/10
  2007-08-06 19:17       ` Nigel Stephens
@ 2007-08-06 19:56         ` Mark Mitchell
  2007-08-06 22:04           ` Fu, Chao-Ying
  0 siblings, 1 reply; 8+ messages in thread
From: Mark Mitchell @ 2007-08-06 19:56 UTC (permalink / raw)
  To: Nigel Stephens; +Cc: Fu, Chao-Ying, gcc-patches, Thekkath, Radhika

Nigel Stephens wrote:
> 
> 
> Mark Mitchell wrote:
>> Fu, Chao-Ying wrote:
>>
>>  
>>> Ex 2: (A and B are negative. )
>>> A.high = 0xffffffffffffffff and A.low = 0x8000000000000000
>>> B.high = 0xffffffffffffffff and B.low = 0x0000000000000001
>>> => A > B, because A.high == B.high and
>>> (unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low
>>>     
>>
>> Yes, I see; two's complement is set up so that after masking the sign
>> bit, you can do an unsigned compare.  Thank you for explaining.
>>   
> 
> Since double_int is already declared with low unsigned, like this:
> 
>    typedef struct
>    {
>      unsigned HOST_WIDE_INT low;
>      HOST_WIDE_INT high;
>    } double_int;
> 
> 
> Would it be more obvious to simply remove the casts altogether, e.g.

There's a comment in double-int.h that suggests that the signedness of
low/high is historical and subject to change.  So, I think the casts are
OK -- but I think omitting them is also OK.  Chao-Ying, you may decide
that as you please.

Thanks,

-- 
Mark Mitchell
CodeSourcery
mark@codesourcery.com
(650) 331-3385 x713

^ permalink raw reply	[flat|nested] 8+ messages in thread

* RE: [ping][patch] Fixed-point patch 2/10
  2007-08-06 19:56         ` Mark Mitchell
@ 2007-08-06 22:04           ` Fu, Chao-Ying
  2007-08-06 23:20             ` Fu, Chao-Ying
  0 siblings, 1 reply; 8+ messages in thread
From: Fu, Chao-Ying @ 2007-08-06 22:04 UTC (permalink / raw)
  To: Mark Mitchell, Stephens, Nigel; +Cc: gcc-patches, Thekkath, Radhika

Mark Mitchell wrote:
> 
> Nigel Stephens wrote:
> > 
> > 
> > Mark Mitchell wrote:
> >> Fu, Chao-Ying wrote:
> >>
> >>  
> >>> Ex 2: (A and B are negative. )
> >>> A.high = 0xffffffffffffffff and A.low = 0x8000000000000000
> >>> B.high = 0xffffffffffffffff and B.low = 0x0000000000000001
> >>> => A > B, because A.high == B.high and
> >>> (unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low
> >>>     
> >>
> >> Yes, I see; two's complement is set up so that after 
> masking the sign
> >> bit, you can do an unsigned compare.  Thank you for explaining.
> >>   
> > 
> > Since double_int is already declared with low unsigned, like this:
> > 
> >    typedef struct
> >    {
> >      unsigned HOST_WIDE_INT low;
> >      HOST_WIDE_INT high;
> >    } double_int;
> > 
> > 
> > Would it be more obvious to simply remove the casts altogether, e.g.
> 
> There's a comment in double-int.h that suggests that the signedness of
> low/high is historical and subject to change.  So, I think 
> the casts are
> OK -- but I think omitting them is also OK.  Chao-Ying, you may decide
> that as you please.
> 
 
  Ok. I will omit them.  Thanks!

Regards,
Chao-ying

^ permalink raw reply	[flat|nested] 8+ messages in thread

* RE: [ping][patch] Fixed-point patch 2/10
  2007-08-06 22:04           ` Fu, Chao-Ying
@ 2007-08-06 23:20             ` Fu, Chao-Ying
  0 siblings, 0 replies; 8+ messages in thread
From: Fu, Chao-Ying @ 2007-08-06 23:20 UTC (permalink / raw)
  To: Mark Mitchell, Stephens, Nigel; +Cc: gcc-patches, Thekkath, Radhika

> 
> Mark Mitchell wrote:
> > 
> > Nigel Stephens wrote:
> > > 
> > > 
> > > Mark Mitchell wrote:
> > >> Fu, Chao-Ying wrote:
> > >>
> > >>  
> > >>> Ex 2: (A and B are negative. )
> > >>> A.high = 0xffffffffffffffff and A.low = 0x8000000000000000
> > >>> B.high = 0xffffffffffffffff and B.low = 0x0000000000000001
> > >>> => A > B, because A.high == B.high and
> > >>> (unsigned HOST_WIDE_INT) a.low > (unsigned HOST_WIDE_INT) b.low
> > >>>     
> > >>
> > >> Yes, I see; two's complement is set up so that after 
> > masking the sign
> > >> bit, you can do an unsigned compare.  Thank you for explaining.
> > >>   
> > > 
> > > Since double_int is already declared with low unsigned, like this:
> > > 
> > >    typedef struct
> > >    {
> > >      unsigned HOST_WIDE_INT low;
> > >      HOST_WIDE_INT high;
> > >    } double_int;
> > > 
> > > 
> > > Would it be more obvious to simply remove the casts 
> altogether, e.g.
> > 
> > There's a comment in double-int.h that suggests that the 
> signedness of
> > low/high is historical and subject to change.  So, I think 
> > the casts are
> > OK -- but I think omitting them is also OK.  Chao-Ying, you 
> may decide
> > that as you please.
> > 
>  
>   Ok. I will omit them.  Thanks!
> 

  This patch (2/10) is applied.  Thanks a lot!

Regards,
Chao-ying

^ permalink raw reply	[flat|nested] 8+ messages in thread

end of thread, other threads:[~2007-08-06 23:20 UTC | newest]

Thread overview: 8+ messages (download: mbox.gz / follow: Atom feed)
-- links below jump to the message on this page --
2007-08-01 23:24 [ping][patch] Fixed-point patch 2/10 Fu, Chao-Ying
2007-08-06  3:50 ` Mark Mitchell
2007-08-06 17:35   ` Fu, Chao-Ying
2007-08-06 18:08     ` Mark Mitchell
2007-08-06 19:17       ` Nigel Stephens
2007-08-06 19:56         ` Mark Mitchell
2007-08-06 22:04           ` Fu, Chao-Ying
2007-08-06 23:20             ` Fu, Chao-Ying

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox;
as well as URLs for read-only IMAP folder(s) and NNTP newsgroup(s).