[Top][All Lists]
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: floor, ceil, trunc, round, minus zero, and alpha
|
From: |
Bruno Haible |
|
Subject: |
Re: floor, ceil, trunc, round, minus zero, and alpha |
|
Date: |
Wed, 22 Dec 2010 16:13:17 +0100 |
|
User-agent: |
KMail/1.9.9 |
Eric Blake wrote:
> Fred Tydeman pointed out that since POSIX [MX] shading defers to C99
> Appendix F, which defers to IEEE 754, and IEEE 754 section 6.3 The Sign
> Bit states:
>
> ... and the sign of the result of the round floating-point number to
> integral value operation is the sign of the operand. These rules shall
> apply even when operands or results are zero or infinite.
>
> Therefore, POSIX requires ceil(-0.5) to return -0.0 when you are
> strictly compliant
Thanks for the clarification. I'm extending the {floor,ceil,trunc,round}*-ieee
modules to require this property:
2010-12-22 Bruno Haible <address@hidden>
round: Implement result sign according to IEEE 754.
* lib/round.c (MIN, MINUS_ZERO): New macros.
(FUNC): Return -0.0 for -0.5 < x < 0.
* tests/test-roundf-ieee.c (main): Test also values between -1 and 1.
* tests/test-round-ieee.c (main): Likewise.
* tests/test-roundl-ieee.c (main): Likewise.
trunc: Implement result sign according to IEEE 754.
* lib/trunc.c (MIN, MINUS_ZERO): New macros.
(FUNC): Return +0.0 for 0 < x < 1 and -0.0 for -1 < x < 0.
* tests/test-trunc2.c: Include minus-zero.h.
(MINUS_ZERO): New macro.
(trunc_reference): Keep in sync with lib/trunc.c.
* tests/test-truncf2.c: Include minus-zero.h.
(MINUS_ZERO): New macro.
(truncf_reference): Keep in sync with lib/trunc.c.
* tests/test-truncf-ieee.c (main): Test also values between -1 and 1.
* tests/test-trunc-ieee.c (main): Likewise.
* tests/test-truncl-ieee.c (main): Likewise.
ceil: Implement result sign according to IEEE 754.
* lib/ceil.c (MIN, MINUS_ZERO): New macros.
(FUNC): Return -0.0 for -1 < x < 0.
* tests/test-ceil2.c: Include minus-zero.h.
(MINUS_ZERO): New macro.
(ceil_reference): Keep in sync with lib/ceil.c.
* tests/test-ceilf2.c: Include minus-zero.h.
(MINUS_ZERO): New macro.
(ceilf_reference): Keep in sync with lib/ceil.c.
* tests/test-ceilf-ieee.c (main): Test also values between -1 and 1.
* tests/test-ceil-ieee.c (main): Likewise.
* tests/test-ceill-ieee.c (main): Likewise.
floor: Implement result sign according to IEEE 754.
* lib/floor.c (FUNC): Return +0.0 for 0 < x < 1.
* tests/test-floor2.c (floor_reference): Keep in sync with lib/floor.c.
* tests/test-floorf2.c (floorf_reference): Likewise.
* tests/test-floorf-ieee.c (main): Test also values between -1 and 1.
* tests/test-floor-ieee.c (main): Likewise.
* tests/test-floorl-ieee.c (main): Likewise.
--- lib/ceil.c.orig Wed Dec 22 15:38:46 2010
+++ lib/ceil.c Wed Dec 22 11:01:28 2010
@@ -1,5 +1,5 @@
/* Round towards positive infinity.
- Copyright (C) 2007, 2009, 2010 Free Software Foundation, Inc.
+ Copyright (C) 2007, 2010 Free Software Foundation, Inc.
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@@ -27,19 +27,29 @@
# define FUNC ceill
# define DOUBLE long double
# define MANT_DIG LDBL_MANT_DIG
+# define MIN LDBL_MIN
# define L_(literal) literal##L
#elif ! defined USE_FLOAT
# define FUNC ceil
# define DOUBLE double
# define MANT_DIG DBL_MANT_DIG
+# define MIN DBL_MIN
# define L_(literal) literal
#else /* defined USE_FLOAT */
# define FUNC ceilf
# define DOUBLE float
# define MANT_DIG FLT_MANT_DIG
+# define MIN FLT_MIN
# define L_(literal) literal##f
#endif
+/* -0.0. See minus-zero.h. */
+#if defined __hpux || defined __sgi || defined __ICC
+# define MINUS_ZERO (-MIN * MIN)
+#else
+# define MINUS_ZERO L_(-0.0)
+#endif
+
/* 2^(MANT_DIG-1). */
static const DOUBLE TWO_MANT_DIG =
/* Assume MANT_DIG <= 5 * 31.
@@ -78,8 +88,12 @@
}
else if (z < L_(0.0))
{
+ /* For -1 < x < 0, return -0.0 regardless of the current rounding
+ mode. */
+ if (z > L_(-1.0))
+ z = MINUS_ZERO;
/* Avoid rounding errors for values near -2^k, where k >= MANT_DIG-1. */
- if (z > - TWO_MANT_DIG)
+ else if (z > - TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z -= TWO_MANT_DIG;
--- lib/floor.c.orig Wed Dec 22 15:38:46 2010
+++ lib/floor.c Wed Dec 22 11:01:02 2010
@@ -1,5 +1,5 @@
/* Round towards negative infinity.
- Copyright (C) 2007, 2009, 2010 Free Software Foundation, Inc.
+ Copyright (C) 2007, 2010 Free Software Foundation, Inc.
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@@ -65,8 +65,12 @@
if (z > L_(0.0))
{
+ /* For 0 < x < 1, return +0.0 even if the current rounding mode is
+ FE_DOWNWARD. */
+ if (z < L_(1.0))
+ z = L_(0.0);
/* Avoid rounding errors for values near 2^k, where k >= MANT_DIG-1. */
- if (z < TWO_MANT_DIG)
+ else if (z < TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z += TWO_MANT_DIG;
--- lib/round.c.orig Wed Dec 22 15:38:46 2010
+++ lib/round.c Wed Dec 22 11:03:16 2010
@@ -1,5 +1,5 @@
/* Round toward nearest, breaking ties away from zero.
- Copyright (C) 2007, 2009, 2010 Free Software Foundation, Inc.
+ Copyright (C) 2007, 2010 Free Software Foundation, Inc.
This program is free software; you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@@ -29,6 +29,7 @@
# define CEIL ceill
# define DOUBLE long double
# define MANT_DIG LDBL_MANT_DIG
+# define MIN LDBL_MIN
# define L_(literal) literal##L
# define HAVE_FLOOR_AND_CEIL HAVE_FLOORL_AND_CEILL
#elif ! defined USE_FLOAT
@@ -37,6 +38,7 @@
# define CEIL ceil
# define DOUBLE double
# define MANT_DIG DBL_MANT_DIG
+# define MIN DBL_MIN
# define L_(literal) literal
# define HAVE_FLOOR_AND_CEIL 1
#else /* defined USE_FLOAT */
@@ -45,10 +47,18 @@
# define CEIL ceilf
# define DOUBLE float
# define MANT_DIG FLT_MANT_DIG
+# define MIN FLT_MIN
# define L_(literal) literal##f
# define HAVE_FLOOR_AND_CEIL HAVE_FLOORF_AND_CEILF
#endif
+/* -0.0. See minus-zero.h. */
+#if defined __hpux || defined __sgi || defined __ICC
+# define MINUS_ZERO (-MIN * MIN)
+#else
+# define MINUS_ZERO L_(-0.0)
+#endif
+
/* If we're being included from test-round2[f].c, it already defined names for
our round implementations. Otherwise, pick the preferred implementation for
this machine. */
@@ -133,7 +143,7 @@
{
/* Avoid rounding error for x = -(0.5 - 2^(-MANT_DIG-1)). */
if (z > - L_(0.5))
- z = L_(0.0);
+ z = MINUS_ZERO;
/* Avoid rounding errors for values near -2^k, where k >= MANT_DIG-1. */
else if (z > -TWO_MANT_DIG)
{
--- lib/trunc.c.orig Wed Dec 22 15:38:46 2010
+++ lib/trunc.c Wed Dec 22 11:02:15 2010
@@ -1,5 +1,5 @@
/* Round towards zero.
- Copyright (C) 2007, 2009, 2010 Free Software Foundation, Inc.
+ Copyright (C) 2007, 2010 Free Software Foundation, Inc.
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
@@ -27,19 +27,29 @@
# define FUNC truncl
# define DOUBLE long double
# define MANT_DIG LDBL_MANT_DIG
+# define MIN LDBL_MIN
# define L_(literal) literal##L
#elif ! defined USE_FLOAT
# define FUNC trunc
# define DOUBLE double
# define MANT_DIG DBL_MANT_DIG
+# define MIN DBL_MIN
# define L_(literal) literal
#else /* defined USE_FLOAT */
# define FUNC truncf
# define DOUBLE float
# define MANT_DIG FLT_MANT_DIG
+# define MIN FLT_MIN
# define L_(literal) literal##f
#endif
+/* -0.0. See minus-zero.h. */
+#if defined __hpux || defined __sgi || defined __ICC
+# define MINUS_ZERO (-MIN * MIN)
+#else
+# define MINUS_ZERO L_(-0.0)
+#endif
+
/* 2^(MANT_DIG-1). */
static const DOUBLE TWO_MANT_DIG =
/* Assume MANT_DIG <= 5 * 31.
@@ -65,8 +75,12 @@
if (z > L_(0.0))
{
+ /* For 0 < x < 1, return +0.0 even if the current rounding mode is
+ FE_DOWNWARD. */
+ if (z < L_(1.0))
+ z = L_(0.0);
/* Avoid rounding errors for values near 2^k, where k >= MANT_DIG-1. */
- if (z < TWO_MANT_DIG)
+ else if (z < TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z += TWO_MANT_DIG;
@@ -78,8 +92,12 @@
}
else if (z < L_(0.0))
{
+ /* For -1 < x < 0, return -0.0 regardless of the current rounding
+ mode. */
+ if (z > L_(-1.0))
+ z = MINUS_ZERO;
/* Avoid rounding errors for values near -2^k, where k >= MANT_DIG-1. */
- if (z > - TWO_MANT_DIG)
+ else if (z > - TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z -= TWO_MANT_DIG;
--- tests/test-ceil-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-ceil-ieee.c Wed Dec 22 09:58:17 2010
@@ -24,9 +24,20 @@
int
main ()
{
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (ceil (0.0)));
ASSERT (!!signbit (ceil (minus_zerod)) == !!signbit (minus_zerod));
+ /* Positive numbers. */
+ ASSERT (!signbit (ceil (0.3)));
+ ASSERT (!signbit (ceil (0.7)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (ceil (-0.3)) == !!signbit (minus_zerod));
+ ASSERT (!!signbit (ceil (-0.7)) == !!signbit (minus_zerod));
return 0;
}
--- tests/test-ceil2.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-ceil2.c Wed Dec 22 11:09:19 2010
@@ -29,6 +29,7 @@
#include <stdio.h>
#include "isnand-nolibm.h"
+#include "minus-zero.h"
#include "macros.h"
@@ -38,6 +39,9 @@
#define MANT_DIG DBL_MANT_DIG
#define L_(literal) literal
+/* -0.0. See minus-zero.h. */
+#define MINUS_ZERO minus_zerod
+
/* 2^(MANT_DIG-1). */
static const DOUBLE TWO_MANT_DIG =
/* Assume MANT_DIG <= 5 * 31.
@@ -79,8 +83,12 @@
}
else if (z < L_(0.0))
{
+ /* For -1 < x < 0, return -0.0 regardless of the current rounding
+ mode. */
+ if (z > L_(-1.0))
+ z = MINUS_ZERO;
/* Avoid rounding errors for values near -2^k, where k >= MANT_DIG-1. */
- if (z > - TWO_MANT_DIG)
+ else if (z > - TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z -= TWO_MANT_DIG;
--- tests/test-ceilf-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-ceilf-ieee.c Wed Dec 22 09:58:17 2010
@@ -37,9 +37,20 @@
{
float (*my_ceilf) (float) = argc ? ceilf : dummy;
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (my_ceilf (0.0f)));
ASSERT (!!signbit (my_ceilf (minus_zerof)) == !!signbit (minus_zerof));
+ /* Positive numbers. */
+ ASSERT (!signbit (my_ceilf (0.3f)));
+ ASSERT (!signbit (my_ceilf (0.7f)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (my_ceilf (-0.3f)) == !!signbit (minus_zerof));
+ ASSERT (!!signbit (my_ceilf (-0.7f)) == !!signbit (minus_zerof));
return 0;
}
--- tests/test-ceilf2.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-ceilf2.c Wed Dec 22 11:09:39 2010
@@ -29,6 +29,7 @@
#include <stdio.h>
#include "isnanf-nolibm.h"
+#include "minus-zero.h"
#include "macros.h"
@@ -38,6 +39,9 @@
#define MANT_DIG FLT_MANT_DIG
#define L_(literal) literal##f
+/* -0.0. See minus-zero.h. */
+#define MINUS_ZERO minus_zerof
+
/* 2^(MANT_DIG-1). */
static const DOUBLE TWO_MANT_DIG =
/* Assume MANT_DIG <= 5 * 31.
@@ -79,8 +83,12 @@
}
else if (z < L_(0.0))
{
+ /* For -1 < x < 0, return -0.0 regardless of the current rounding
+ mode. */
+ if (z > L_(-1.0))
+ z = MINUS_ZERO;
/* Avoid rounding errors for values near -2^k, where k >= MANT_DIG-1. */
- if (z > - TWO_MANT_DIG)
+ else if (z > - TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z -= TWO_MANT_DIG;
--- tests/test-ceill-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-ceill-ieee.c Wed Dec 22 09:58:17 2010
@@ -29,9 +29,20 @@
BEGIN_LONG_DOUBLE_ROUNDING ();
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (ceill (0.0L)));
ASSERT (!!signbit (ceill (minus_zerol)) == !!signbit (minus_zerol));
+ /* Positive numbers. */
+ ASSERT (!signbit (ceill (0.3L)));
+ ASSERT (!signbit (ceill (0.7L)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (ceill (-0.3L)) == !!signbit (minus_zerol));
+ ASSERT (!!signbit (ceill (-0.7L)) == !!signbit (minus_zerol));
return 0;
}
--- tests/test-floor-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-floor-ieee.c Wed Dec 22 09:58:18 2010
@@ -24,9 +24,20 @@
int
main ()
{
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (floor (0.0)));
ASSERT (!!signbit (floor (minus_zerod)) == !!signbit (minus_zerod));
+ /* Positive numbers. */
+ ASSERT (!signbit (floor (0.3)));
+ ASSERT (!signbit (floor (0.7)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (floor (-0.3)) == !!signbit (minus_zerod));
+ ASSERT (!!signbit (floor (-0.7)) == !!signbit (minus_zerod));
return 0;
}
--- tests/test-floor2.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-floor2.c Wed Dec 22 11:07:16 2010
@@ -63,8 +63,12 @@
if (z > L_(0.0))
{
+ /* For 0 < x < 1, return +0.0 even if the current rounding mode is
+ FE_DOWNWARD. */
+ if (z < L_(1.0))
+ z = L_(0.0);
/* Avoid rounding errors for values near 2^k, where k >= MANT_DIG-1. */
- if (z < TWO_MANT_DIG)
+ else if (z < TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z += TWO_MANT_DIG;
--- tests/test-floorf-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-floorf-ieee.c Wed Dec 22 09:58:17 2010
@@ -37,9 +37,20 @@
{
float (*my_floorf) (float) = argc ? floorf : dummy;
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (my_floorf (0.0f)));
ASSERT (!!signbit (my_floorf (minus_zerof)) == !!signbit (minus_zerof));
+ /* Positive numbers. */
+ ASSERT (!signbit (my_floorf (0.3f)));
+ ASSERT (!signbit (my_floorf (0.7f)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (my_floorf (-0.3f)) == !!signbit (minus_zerof));
+ ASSERT (!!signbit (my_floorf (-0.7f)) == !!signbit (minus_zerof));
return 0;
}
--- tests/test-floorf2.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-floorf2.c Wed Dec 22 11:06:23 2010
@@ -63,8 +63,12 @@
if (z > L_(0.0))
{
+ /* For 0 < x < 1, return +0.0 even if the current rounding mode is
+ FE_DOWNWARD. */
+ if (z < L_(1.0))
+ z = L_(0.0);
/* Avoid rounding errors for values near 2^k, where k >= MANT_DIG-1. */
- if (z < TWO_MANT_DIG)
+ else if (z < TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z += TWO_MANT_DIG;
--- tests/test-floorl-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-floorl-ieee.c Wed Dec 22 09:58:18 2010
@@ -29,9 +29,20 @@
BEGIN_LONG_DOUBLE_ROUNDING ();
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (floorl (0.0L)));
ASSERT (!!signbit (floorl (minus_zerol)) == !!signbit (minus_zerol));
+ /* Positive numbers. */
+ ASSERT (!signbit (floorl (0.3L)));
+ ASSERT (!signbit (floorl (0.7L)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (floorl (-0.3L)) == !!signbit (minus_zerol));
+ ASSERT (!!signbit (floorl (-0.7L)) == !!signbit (minus_zerol));
return 0;
}
--- tests/test-round-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-round-ieee.c Wed Dec 22 09:58:18 2010
@@ -25,9 +25,20 @@
int
main ()
{
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (round (0.0)));
ASSERT (!!signbit (round (minus_zerod)) == !!signbit (minus_zerod));
+ /* Positive numbers. */
+ ASSERT (!signbit (round (0.3)));
+ ASSERT (!signbit (round (0.7)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (round (-0.3)) == !!signbit (minus_zerod));
+ ASSERT (!!signbit (round (-0.7)) == !!signbit (minus_zerod));
return 0;
}
--- tests/test-roundf-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-roundf-ieee.c Wed Dec 22 09:58:18 2010
@@ -25,9 +25,20 @@
int
main ()
{
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (roundf (0.0f)));
ASSERT (!!signbit (roundf (minus_zerof)) == !!signbit (minus_zerof));
+ /* Positive numbers. */
+ ASSERT (!signbit (roundf (0.3f)));
+ ASSERT (!signbit (roundf (0.7f)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (roundf (-0.3f)) == !!signbit (minus_zerof));
+ ASSERT (!!signbit (roundf (-0.7f)) == !!signbit (minus_zerof));
return 0;
}
--- tests/test-roundl-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-roundl-ieee.c Wed Dec 22 09:59:25 2010
@@ -30,9 +30,20 @@
BEGIN_LONG_DOUBLE_ROUNDING ();
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (roundl (0.0L)));
ASSERT (!!signbit (roundl (minus_zerol)) == !!signbit (minus_zerol));
+ /* Positive numbers. */
+ ASSERT (!signbit (roundl (0.3L)));
+ ASSERT (!signbit (roundl (0.7L)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (roundl (-0.3L)) == !!signbit (minus_zerol));
+ ASSERT (!!signbit (roundl (-0.7L)) == !!signbit (minus_zerol));
return 0;
}
--- tests/test-trunc-ieee.c.orig Wed Dec 22 15:38:46 2010
+++ tests/test-trunc-ieee.c Wed Dec 22 09:58:18 2010
@@ -24,9 +24,20 @@
int
main ()
{
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (trunc (0.0)));
ASSERT (!!signbit (trunc (minus_zerod)) == !!signbit (minus_zerod));
+ /* Positive numbers. */
+ ASSERT (!signbit (trunc (0.3)));
+ ASSERT (!signbit (trunc (0.7)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (trunc (-0.3)) == !!signbit (minus_zerod));
+ ASSERT (!!signbit (trunc (-0.7)) == !!signbit (minus_zerod));
return 0;
}
--- tests/test-trunc2.c.orig Wed Dec 22 15:38:47 2010
+++ tests/test-trunc2.c Wed Dec 22 11:10:26 2010
@@ -29,6 +29,7 @@
#include <stdio.h>
#include "isnand-nolibm.h"
+#include "minus-zero.h"
#include "macros.h"
@@ -38,6 +39,9 @@
#define MANT_DIG DBL_MANT_DIG
#define L_(literal) literal
+/* -0.0. See minus-zero.h. */
+#define MINUS_ZERO minus_zerod
+
/* 2^(MANT_DIG-1). */
static const DOUBLE TWO_MANT_DIG =
/* Assume MANT_DIG <= 5 * 31.
@@ -63,8 +67,12 @@
if (z > L_(0.0))
{
+ /* For 0 < x < 1, return +0.0 even if the current rounding mode is
+ FE_DOWNWARD. */
+ if (z < L_(1.0))
+ z = L_(0.0);
/* Avoid rounding errors for values near 2^k, where k >= MANT_DIG-1. */
- if (z < TWO_MANT_DIG)
+ else if (z < TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z += TWO_MANT_DIG;
@@ -76,8 +84,12 @@
}
else if (z < L_(0.0))
{
+ /* For -1 < x < 0, return -0.0 regardless of the current rounding
+ mode. */
+ if (z > L_(-1.0))
+ z = MINUS_ZERO;
/* Avoid rounding errors for values near -2^k, where k >= MANT_DIG-1. */
- if (z > - TWO_MANT_DIG)
+ else if (z > - TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z -= TWO_MANT_DIG;
--- tests/test-truncf-ieee.c.orig Wed Dec 22 15:38:47 2010
+++ tests/test-truncf-ieee.c Wed Dec 22 09:58:18 2010
@@ -24,9 +24,20 @@
int
main ()
{
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (truncf (0.0f)));
ASSERT (!!signbit (truncf (minus_zerof)) == !!signbit (minus_zerof));
+ /* Positive numbers. */
+ ASSERT (!signbit (truncf (0.3f)));
+ ASSERT (!signbit (truncf (0.7f)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (truncf (-0.3f)) == !!signbit (minus_zerof));
+ ASSERT (!!signbit (truncf (-0.7f)) == !!signbit (minus_zerof));
return 0;
}
--- tests/test-truncf2.c.orig Wed Dec 22 15:38:47 2010
+++ tests/test-truncf2.c Wed Dec 22 11:11:03 2010
@@ -29,6 +29,7 @@
#include <stdio.h>
#include "isnanf-nolibm.h"
+#include "minus-zero.h"
#include "macros.h"
@@ -38,6 +39,9 @@
#define MANT_DIG FLT_MANT_DIG
#define L_(literal) literal##f
+/* -0.0. See minus-zero.h. */
+#define MINUS_ZERO minus_zerof
+
/* 2^(MANT_DIG-1). */
static const DOUBLE TWO_MANT_DIG =
/* Assume MANT_DIG <= 5 * 31.
@@ -63,8 +67,12 @@
if (z > L_(0.0))
{
+ /* For 0 < x < 1, return +0.0 even if the current rounding mode is
+ FE_DOWNWARD. */
+ if (z < L_(1.0))
+ z = L_(0.0);
/* Avoid rounding errors for values near 2^k, where k >= MANT_DIG-1. */
- if (z < TWO_MANT_DIG)
+ else if (z < TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z += TWO_MANT_DIG;
@@ -76,8 +84,12 @@
}
else if (z < L_(0.0))
{
+ /* For -1 < x < 0, return -0.0 regardless of the current rounding
+ mode. */
+ if (z > L_(-1.0))
+ z = MINUS_ZERO;
/* Avoid rounding errors for values near -2^k, where k >= MANT_DIG-1. */
- if (z > - TWO_MANT_DIG)
+ else if (z > - TWO_MANT_DIG)
{
/* Round to the next integer (nearest or up or down, doesn't
matter). */
z -= TWO_MANT_DIG;
--- tests/test-truncl-ieee.c.orig Wed Dec 22 15:38:47 2010
+++ tests/test-truncl-ieee.c Wed Dec 22 09:58:18 2010
@@ -29,9 +29,20 @@
BEGIN_LONG_DOUBLE_ROUNDING ();
+ /* See IEEE 754, section 6.3:
+ "the sign of the result of the round floating-point number to
+ integral value operation is the sign of the operand. These rules
+ shall apply even when operands or results are zero or infinite." */
+
/* Zero. */
ASSERT (!signbit (truncl (0.0L)));
ASSERT (!!signbit (truncl (minus_zerol)) == !!signbit (minus_zerol));
+ /* Positive numbers. */
+ ASSERT (!signbit (truncl (0.3L)));
+ ASSERT (!signbit (truncl (0.7L)));
+ /* Negative numbers. */
+ ASSERT (!!signbit (truncl (-0.3L)) == !!signbit (minus_zerol));
+ ASSERT (!!signbit (truncl (-0.7L)) == !!signbit (minus_zerol));
return 0;
}
- floor, ceil, trunc, round, minus zero, and alpha, Bruno Haible, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha, Eric Blake, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha, Bruno Haible, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha, Eric Blake, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha, Bruno Haible, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha, Paul Eggert, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha, Bruno Haible, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha, Bruno Haible, 2010/12/22
- Re: floor, ceil, trunc, round, minus zero, and alpha, Eric Blake, 2010/12/20
- Re: floor, ceil, trunc, round, minus zero, and alpha,
Bruno Haible <=